diff --git a/.gitignore b/.gitignore
index 9622ab78e0..4f21fefda9 100644
--- a/.gitignore
+++ b/.gitignore
@@ -22,6 +22,7 @@ cmake-build-*
 
 # generated while compiling
 python/paddle/v2/framework/core.so
+paddle/pybind/pybind.h
 CMakeFiles
 cmake_install.cmake
 paddle/.timestamp
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 08237cd850..5739c2a260 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -67,6 +67,9 @@ endif()
 if(ANDROID)
     if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "16")
         message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 16")
+    elseif(${CMAKE_SYSTEM_VERSION} VERSION_LESS "21")
+        # TODO: support glog for Android api 16 ~ 19 in the future
+        message(WARNING "Using the unofficial git repository <https://github.com/Xreki/glog.git> instead")
     endif()
 
     set(WITH_GPU OFF CACHE STRING
diff --git a/Dockerfile.android b/Dockerfile.android
index 452aa15745..9d13a414f6 100644
--- a/Dockerfile.android
+++ b/Dockerfile.android
@@ -6,13 +6,14 @@ RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ub
 
 # ENV variables
 ARG ANDROID_ABI
+ARG ANDROID_API
 
 ENV ANDROID_ABI=${ANDROID_ABI:-"armeabi-v7a"}
+ENV ANDROID_API=${ANDROID_API:-21}
 
 ENV HOME=/root \
     ANDROID_NDK_HOME=/opt/android-ndk-linux \
-    ANDROID_ARM_STANDALONE_TOOLCHAIN=/opt/arm-toolchain \
-    ANDROID_ARM64_STANDALONE_TOOLCHAIN=/opt/arm64-toolchain
+    ANDROID_TOOLCHAINS_DIR=/opt/toolchains
 
 RUN apt-get update && \
     apt-get install -y \
@@ -42,14 +43,12 @@ RUN pip install --upgrade pip && \
     pip install pre-commit
 
 # Android NDK
-RUN mkdir /opt/android-ndk-tmp && \
+RUN mkdir -p ${ANDROID_TOOLCHAINS_DIR} && \
+    mkdir -p /opt/android-ndk-tmp && \
     cd /opt/android-ndk-tmp && \
     wget -q https://dl.google.com/android/repository/android-ndk-r14b-linux-x86_64.zip && \
     unzip -q android-ndk-r14b-linux-x86_64.zip && \
     mv android-ndk-r14b ${ANDROID_NDK_HOME} && \
-    ${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh --arch=arm --platform=android-23 --install-dir=${ANDROID_ARM_STANDALONE_TOOLCHAIN} && \
-    ${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh --arch=arm64 --platform=android-23 --install-dir=${ANDROID_ARM64_STANDALONE_TOOLCHAIN} && \
-    rm -rf /opt/android-ndk-tmp && \
-    rm -rf ${ANDROID_NDK_HOME}
+    rm -rf /opt/android-ndk-tmp
 
 CMD ["bash", "/paddle/paddle/scripts/docker/build_android.sh"]
diff --git a/cmake/cpplint.cmake b/cmake/cpplint.cmake
index 8d5d533126..4823dc3e91 100644
--- a/cmake/cpplint.cmake
+++ b/cmake/cpplint.cmake
@@ -26,9 +26,9 @@ set(IGNORE_PATTERN
     .*ImportanceSampler.*
     .*cblas\\.h.*
     .*\\.pb\\.txt
-    .*LtrDataProvider.*
     .*MultiDataProvider.*
-    .*pb.*)
+    .*pb.*
+    .*pybind.h)
 
 # add_style_check_target
 #
diff --git a/cmake/external/gflags.cmake b/cmake/external/gflags.cmake
index 16e5bef4cd..01a2f4d5fa 100644
--- a/cmake/external/gflags.cmake
+++ b/cmake/external/gflags.cmake
@@ -18,9 +18,9 @@ SET(GFLAGS_SOURCES_DIR ${THIRD_PARTY_PATH}/gflags)
 SET(GFLAGS_INSTALL_DIR ${THIRD_PARTY_PATH}/install/gflags)
 SET(GFLAGS_INCLUDE_DIR "${GFLAGS_INSTALL_DIR}/include" CACHE PATH "gflags include directory." FORCE)
 IF(WIN32)
-    set(GFLAGS_LIBRARIES "${GFLAGS_INSTALL_DIR}/lib/gflags.lib" CACHE FILEPATH "GFLAGS_LIBRARIES" FORCE)
+  set(GFLAGS_LIBRARIES "${GFLAGS_INSTALL_DIR}/lib/gflags.lib" CACHE FILEPATH "GFLAGS_LIBRARIES" FORCE)
 ELSE(WIN32)
-    set(GFLAGS_LIBRARIES "${GFLAGS_INSTALL_DIR}/lib/libgflags.a" CACHE FILEPATH "GFLAGS_LIBRARIES" FORCE)
+  set(GFLAGS_LIBRARIES "${GFLAGS_INSTALL_DIR}/lib/libgflags.a" CACHE FILEPATH "GFLAGS_LIBRARIES" FORCE)
 ENDIF(WIN32)
 
 INCLUDE_DIRECTORIES(${GFLAGS_INCLUDE_DIR})
@@ -56,3 +56,12 @@ SET_PROPERTY(TARGET gflags PROPERTY IMPORTED_LOCATION ${GFLAGS_LIBRARIES})
 ADD_DEPENDENCIES(gflags extern_gflags)
 
 LIST(APPEND external_project_dependencies gflags)
+
+IF(WITH_C_API)
+  INSTALL(DIRECTORY ${GFLAGS_INCLUDE_DIR} DESTINATION third_party/gflags)
+  IF(ANDROID)
+    INSTALL(FILES ${GFLAGS_LIBRARIES} DESTINATION third_party/gflags/lib/${ANDROID_ABI})
+  ELSE()
+    INSTALL(FILES ${GFLAGS_LIBRARIES} DESTINATION third_party/gflags/lib)
+  ENDIF()
+ENDIF()
diff --git a/cmake/external/glog.cmake b/cmake/external/glog.cmake
index 8a594a825a..b450a30166 100644
--- a/cmake/external/glog.cmake
+++ b/cmake/external/glog.cmake
@@ -19,9 +19,9 @@ SET(GLOG_INSTALL_DIR ${THIRD_PARTY_PATH}/install/glog)
 SET(GLOG_INCLUDE_DIR "${GLOG_INSTALL_DIR}/include" CACHE PATH "glog include directory." FORCE)
 
 IF(WIN32)
-    SET(GLOG_LIBRARIES "${GLOG_INSTALL_DIR}/lib/libglog.lib" CACHE FILEPATH "glog library." FORCE)
+  SET(GLOG_LIBRARIES "${GLOG_INSTALL_DIR}/lib/libglog.lib" CACHE FILEPATH "glog library." FORCE)
 ELSE(WIN32)
-    SET(GLOG_LIBRARIES "${GLOG_INSTALL_DIR}/lib/libglog.a" CACHE FILEPATH "glog library." FORCE)
+  SET(GLOG_LIBRARIES "${GLOG_INSTALL_DIR}/lib/libglog.a" CACHE FILEPATH "glog library." FORCE)
 ENDIF(WIN32)
 
 INCLUDE_DIRECTORIES(${GLOG_INCLUDE_DIR})
@@ -56,3 +56,12 @@ ADD_DEPENDENCIES(glog extern_glog gflags)
 LINK_LIBRARIES(glog gflags)
 
 LIST(APPEND external_project_dependencies glog)
+
+IF(WITH_C_API)
+  INSTALL(DIRECTORY ${GLOG_INCLUDE_DIR} DESTINATION third_party/glog)
+  IF(ANDROID)
+    INSTALL(FILES ${GLOG_LIBRARIES} DESTINATION third_party/glog/lib/${ANDROID_ABI})
+  ELSE()
+    INSTALL(FILES ${GLOG_LIBRARIES} DESTINATION third_party/glog/lib)
+  ENDIF()
+ENDIF()
diff --git a/cmake/external/openblas.cmake b/cmake/external/openblas.cmake
index f9e05af59f..4fc8d43fc1 100644
--- a/cmake/external/openblas.cmake
+++ b/cmake/external/openblas.cmake
@@ -73,6 +73,26 @@ IF(NOT ${CBLAS_FOUND})
         UPDATE_COMMAND      ""
         CONFIGURE_COMMAND   ""
     )
+
+    IF(WITH_C_API)
+        INSTALL(DIRECTORY ${CBLAS_INC_DIR} DESTINATION third_party/openblas)
+        # Because libopenblas.a is a symbolic link of another library, thus need to
+        # install the whole directory.
+        IF(ANDROID)
+            SET(TMP_INSTALL_DIR third_party/openblas/lib/${ANDROID_ABI})
+        ELSE()
+            SET(TMP_INSTALL_DIR third_party/openblas/lib)
+        ENDIF()
+        INSTALL(CODE "execute_process(
+            COMMAND ${CMAKE_COMMAND} -E copy_directory ${CBLAS_INSTALL_DIR}/lib
+                    destination ${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}
+            )"
+        )
+        INSTALL(CODE "MESSAGE(STATUS \"Installing: \"
+                \"${CBLAS_INSTALL_DIR}/lib -> ${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}\"
+            )"
+        )
+    ENDIF()
 ENDIF(NOT ${CBLAS_FOUND})
 
 MESSAGE(STATUS "BLAS library: ${CBLAS_LIBRARIES}")
diff --git a/cmake/external/protobuf.cmake b/cmake/external/protobuf.cmake
index e629d61585..a887be2e2a 100644
--- a/cmake/external/protobuf.cmake
+++ b/cmake/external/protobuf.cmake
@@ -223,6 +223,15 @@ IF(NOT PROTOBUF_FOUND)
     SET(PROTOBUF_PROTOC_LIBRARY ${extern_protobuf_PROTOC_LIBRARY}
         CACHE FILEPATH "protoc library." FORCE)
 
+    IF(WITH_C_API)
+        INSTALL(DIRECTORY ${PROTOBUF_INCLUDE_DIR} DESTINATION third_party/protobuf)
+        IF(ANDROID)
+            INSTALL(FILES ${PROTOBUF_LIBRARY} DESTINATION third_party/protobuf/lib/${ANDROID_ABI})
+        ELSE()
+            INSTALL(FILES ${PROTOBUF_LIBRARY} DESTINATION third_party/protobuf/lib)
+        ENDIF()
+    ENDIF()
+
     IF(CMAKE_CROSSCOMPILING)
         PROMPT_PROTOBUF_LIB(protobuf_host extern_protobuf)
     ELSE()
diff --git a/cmake/external/zlib.cmake b/cmake/external/zlib.cmake
index 45ca5542b7..5aecab90ca 100644
--- a/cmake/external/zlib.cmake
+++ b/cmake/external/zlib.cmake
@@ -49,3 +49,12 @@ ExternalProject_Add(
 )
 
 LIST(APPEND external_project_dependencies zlib)
+
+IF(WITH_C_API)
+  INSTALL(DIRECTORY ${ZLIB_INCLUDE_DIR} DESTINATION third_party/zlib)
+  IF(ANDROID)
+    INSTALL(FILES ${ZLIB_LIBRARIES} DESTINATION third_party/zlib/lib/${ANDROID_ABI})
+  ELSE()
+    INSTALL(FILES ${ZLIB_LIBRARIES} DESTINATION third_party/zlib/lib)
+  ENDIF()
+ENDIF()
diff --git a/doc/design/block.md b/doc/design/block.md
new file mode 100644
index 0000000000..be88001220
--- /dev/null
+++ b/doc/design/block.md
@@ -0,0 +1,338 @@
+# Design Doc: Block and Scope
+
+## The Representation of Computation
+
+Both deep learning systems and programming languages help users describe computation procedures.  These systems use various representations of computation:
+
+- Caffe, Torch, and Paddle: sequences of layers.
+- TensorFlow, Caffe2, Mxnet: graphs of operators.
+- PaddlePaddle: nested blocks, like C++ and Java programs.
+
+## Block in Programming Languages and Deep Learning
+
+In programming languages, a block is a pair of curly braces that includes local variables definitions and a sequence of instructions, or operators.
+
+Blocks work with control flow structures like `if`, `else`, and `for`, which have equivalents in deep learning:
+
+| programming languages | PaddlePaddle          |
+|-----------------------|-----------------------|
+| for, while loop       | RNN, WhileOp          |
+| if, if-else, switch   | IfElseOp, SwitchOp    |
+| sequential execution  | a sequence of layers  |
+
+A key difference is that a C++ program describes a one pass computation, whereas a deep learning program describes both the forward and backward passes.
+
+## Stack Frames and the Scope Hierarchy
+
+The existence of the backward makes the execution of a block of traditional programs and PaddlePaddle different to each other:
+
+| programming languages | PaddlePaddle                  |
+|-----------------------|-------------------------------|
+| stack                 | scope hierarchy               |
+| stack frame           | scope                         |
+| push at entering block| push at entering block        |
+| pop at leaving block  | destroy at minibatch completes|
+
+1. In traditional programs:
+
+   - When the execution enters the left curly brace of a block, the runtime pushes a frame into the stack, where it realizes local variables.
+   - After the execution leaves the right curly brace, the runtime pops the frame.
+   - The maximum number of frames in the stack is the maximum depth of nested blocks.
+
+1. In PaddlePaddle
+
+   - When the execution enters a block, PaddlePaddle adds a new scope, where it realizes variables.
+   - PaddlePaddle doesn't pop a scope after the execution of the block because variables therein are to be used by the backward pass.  So it has a stack forest known as a *scope hierarchy*.
+   - The height of the highest tree is the maximum depth of nested blocks.
+   - After the process of a minibatch, PaddlePaddle destroys the scope hierarchy.
+
+## Use Blocks in C++ and PaddlePaddle Programs
+
+Let us consolidate the discussion by presenting some examples.
+
+### Blocks with `if-else` and `IfElseOp`
+
+The following C++ programs shows how blocks are used with the `if-else` structure:
+
+```c++
+int x = 10;
+int y = 20;
+int out;
+bool cond = false;
+if (cond) {
+  int z = x + y;
+  out = softmax(z);
+} else {
+  int z = fc(x);
+  out = z;
+}
+```
+
+An equivalent PaddlePaddle program from the design doc of the [IfElseOp operator](./if_else_op.md) is as follows:
+
+```python
+import paddle as pd
+
+x = var(10)
+y = var(20)
+cond = var(false)
+ie = pd.create_ifelseop(inputs=[x], output_num=1)
+with ie.true_block():
+    x = ie.inputs(true, 0)
+    z = operator.add(x, y)
+    ie.set_output(true, 0, operator.softmax(z))
+with ie.false_block():
+    x = ie.inputs(false, 0)
+    z = layer.fc(x)
+    ie.set_output(true, 0, operator.softmax(z))
+out = b(cond)
+```
+
+In both examples, the left branch computes `softmax(x+y)` and the right branch computes `fc(x)`.
+
+A difference is that variables in the C++ program contain scalar values, whereas those in the PaddlePaddle programs are mini-batches of instances.  The `ie.input(true, 0)` invocation returns instances in the 0-th input, `x`, that corresponds to true values in `cond` as the local variable `x`, where `ie.input(false, 0)` returns instances corresponding to false values.
+
+### Blocks with `for` and `RNNOp`
+
+The following RNN model from the [RNN design doc](./rnn.md)
+
+```python
+x = sequence([10, 20, 30])
+m = var(0)
+W = tensor()
+U = tensor()
+
+rnn = create_rnn(inputs=[input])
+with rnn.stepnet() as net:
+  x = net.set_inputs(0)
+  h = net.add_memory(init=m)
+  fc_out = pd.matmul(W, x)
+  hidden_out = pd.matmul(U, h.pre(n=1))
+  sum = pd.add_two(fc_out, hidden_out)
+  act = pd.sigmoid(sum)
+  h.update(act)                       # update memory with act
+  net.set_outputs(0, act, hidden_out) # two outputs
+
+o1, o2 = rnn()
+print o1, o2
+```
+
+has its equivalent C++ program as follows
+
+```c++
+int* x = {10, 20, 30};
+int m = 0;
+int W = some_value();
+int U = some_other_value();
+
+int mem[sizeof(x) / sizeof(x[0]) + 1];
+int o1[sizeof(x) / sizeof(x[0]) + 1];
+int o2[sizeof(x) / sizeof(x[0]) + 1];
+for (int i = 1; i <= sizeof(x)/sizeof(x[0]); ++i) {
+  int x = x[i-1];
+  if (i == 1) mem[0] = m;
+  int fc_out = W * x;
+  int hidden_out = Y * mem[i-1];
+  int sum = fc_out + hidden_out;
+  int act = sigmoid(sum);
+  mem[i] = act;
+  o1[i] = act;
+  o2[i] = hidden_out;
+}
+
+print_array(o1);
+print_array(o2);
+```
+
+
+## Compilation and Execution
+
+Like TensorFlow programs, a PaddlePaddle program is written in Python.  The first part describes a neural network as a protobuf message, and the rest part executes the message for training or inference.
+
+The generation of this protobuf message is like what a compiler generates a binary executable file.  The execution of the message that the OS executes the binary file.
+
+## The "Binary Executable File Format"
+
+The definition of the protobuf message is as follows:
+
+```protobuf
+message BlockDesc {
+  repeated VarDesc vars = 1;
+  repeated OpDesc ops = 2;
+}
+```
+
+The step net in above RNN example would look like
+
+```
+BlockDesc {
+  vars = {
+    VarDesc {...} // x
+    VarDesc {...} // h
+    VarDesc {...} // fc_out
+    VarDesc {...} // hidden_out
+    VarDesc {...} // sum
+    VarDesc {...} // act
+  }
+  ops = {
+    OpDesc {...} // matmul
+    OpDesc {...} // add_two
+    OpDesc {...} // sigmoid
+  }
+};
+```
+
+Also, the RNN operator in above example is serialized into a protobuf message of type `OpDesc` and would look like:
+
+```
+OpDesc {
+  inputs = {0} // the index of x
+  outputs = {5, 3} // indices of act and hidden_out
+  attrs {
+    "memories" : {1} // the index of h
+    "step_net" : <above step net>
+  }
+};
+```
+
+This `OpDesc` value is in the `ops` field of the `BlockDesc` value representing the global block.
+
+
+## The Compilation of Blocks
+
+During the generation of the Protobuf message, the Block should store VarDesc (the Protobuf message which describes Variable) and OpDesc (the Protobuf message which describes Operator).
+
+VarDesc in a block should have its name scope to avoid local variables affect parent block's name scope.
+Child block's name scopes should inherit the parent's so that OpDesc in child block can reference a VarDesc that stored in parent block. For example
+
+```python
+a = pd.Varaible(shape=[20, 20])
+b = pd.fc(a, params=["fc.w", "fc.b"])
+
+rnn = pd.create_rnn()
+with rnn.stepnet() as net:
+    x = net.set_inputs(a)
+    # reuse fc's parameter
+    fc_without_b = pd.get_variable("fc.w")
+    net.set_outputs(fc_without_b)
+
+out = rnn()
+```
+the method `pd.get_variable` can help retrieve a Variable by a name, a Variable may store in a parent block, but might be retrieved in a child block, so block should have a variable scope that supports inheritance.
+
+In compiler design, the symbol table is a data structure created and maintained by compilers to store information about the occurrence of various entities such as variable names, function names, classes, etc.
+
+To store the definition of variables and operators, we define a C++ class `SymbolTable`, like the one used in compilers.
+
+`SymbolTable` can do the following stuff:
+
+- store the definitions (some names and attributes) of variables and operators,
+- to verify if a variable was declared,
+- to make it possible to implement type checking (offer Protobuf message pointers to `InferShape` handlers).
+
+
+```c++
+// Information in SymbolTable is enough to trace the dependency graph. So maybe
+// the Eval() interface takes a SymbolTable is enough.
+class SymbolTable {
+ public:
+  SymbolTable(SymbolTable* parent) : parent_(parent) {}
+
+  OpDesc* NewOp(const string& name="");
+
+  // TODO determine whether name is generated by python or C++
+  // currently assume that a unique name will be generated by C++ if the
+  // argument name left default.
+  VarDesc* NewVar(const string& name="");
+
+  // find a VarDesc by name, if recursive true, find parent's SymbolTable
+  // recursively.
+  // this interface is introduced to support InferShape, find protobuf messages
+  // of variables and operators, pass pointers into InferShape.
+  // operator
+  //
+  // NOTE maybe some C++ classes such as VarDescBuilder and OpDescBuilder should
+  // be proposed and embedded into pybind to enable python operate on C++ pointers.
+  VarDesc* FindVar(const string& name, bool recursive=true);
+
+  OpDesc* FindOp(const string& name);
+
+  BlockDesc Compile() const;
+
+ private:
+  SymbolTable* parent_;
+
+  map<string, OpDesc> ops_;
+  map<string, VarDesc> vars_;
+};
+```
+
+After all the description of variables and operators is added into SymbolTable,
+the block has enough information to run.
+
+The `Block` class takes a `BlockDesc` as input, and provide `Run` and `InferShape` functions.
+
+
+```c++
+namespace {
+
+class Block : OperatorBase {
+public:
+  Block(const BlockDesc& desc) desc_(desc) {}
+
+  void InferShape(const framework::Scope& scope) const override {
+    if (!symbols_ready_) {
+      CreateVariables(scope);
+      CreateOperators();
+    }
+    // should run InferShape first.
+    for (auto& op : runtime_table_.ops()) {
+      op->InferShape(scope);
+    }
+  }
+
+  void Run(const framework::Scope& scope,
+           const platform::DeviceContext& dev_ctx) const override {
+    PADDLE_ENFORCE(symbols_ready_, "operators and variables should be created first.");
+    for (auto& op : runtime_table_.ops()) {
+      op->Run(scope, dev_ctx);
+    }
+  }
+
+  void CreateVariables(const framework::Scope& scope);
+  void CreateOperators();
+
+  // some other necessary interfaces of NetOp are list below
+  // ...
+
+private:
+  BlockDesc desc_;
+  bool symbols_ready_{false};
+};
+```
+
+## The Execution of Blocks
+
+Block inherits from OperatorBase, which has a Run method.
+Block's Run method will run its operators sequentially.
+
+There is another important interface called `Eval`, which take some arguments called targets, and generate a minimal graph which takes targets as the end points and creates a new Block,
+after `Run`, `Eval` will get the latest value and return the targets.
+
+The definition of Eval is as follows:
+
+```c++
+// clean a block description by targets using the corresponding dependency graph.
+// return a new BlockDesc with minimal number of operators.
+// NOTE not return a Block but the block's description so that this can be distributed
+// to a cluster.
+BlockDesc Prune(const BlockDesc& desc, vector<string> targets);
+
+void Block::Eval(const vector<string>& targets,
+                 const framework::Scope& scope,
+                 const platform::DeviceContext& dev_ctx) {
+  BlockDesc min_desc = Prune(desc_, targets);
+  Block min_block(min_desc);
+  min_block.Run(scope, dev_ctx);
+}
+```
diff --git a/doc/design/if_else_op.md b/doc/design/if_else_op.md
index 7370c2a24f..954a19c073 100644
--- a/doc/design/if_else_op.md
+++ b/doc/design/if_else_op.md
@@ -1,22 +1,4 @@
-IfOp should have only one branch. An IfOp operator takes a `cond` variable whose value must be a vector of N boolean elements. Its return value has M (M<=N) instances, each corresponds to a true element in `cond`.
-
-```python
-import paddle as pd
-
-x = var()
-y = var()
-cond = var()
-
-b = pd.create_ifop(inputs=[x], output_num=1)
-with b.true_block():
-    x = b.inputs(0)
-    z = operator.add(x, y)
-    b.set_output(0, operator.softmax(z))
-
-out = b(cond)
-```
-
-If we want the output still has N instances, we can use IfElseOp with a default value, whose minibatch size must be N:
+IfOp should have only one branch. An IfOp operator takes a `cond` variable whose value must be a vector of N boolean elements. Its return value has N instances. If cond[i] == True, input instance input[i] will go through true_block() and generate output[i]; otherwise it will produce output from false_bloack().
 
 ```python
 import paddle as pd
@@ -39,7 +21,7 @@ with b.false_block():
 out = b(cond)
 ```
 
-If only true_block is set in an IfElseOp, we can have a default value for false as:
+If only true_block is set in an IfElseOp, a special case is that we can have a default value for false as:
 ```python
 import paddle as pd
 
diff --git a/doc/design/ops/images/2_level_rnn.dot b/doc/design/ops/images/2_level_rnn.dot
new file mode 100644
index 0000000000..a498e882a3
--- /dev/null
+++ b/doc/design/ops/images/2_level_rnn.dot
@@ -0,0 +1,56 @@
+digraph G {
+
+  rnn [label="1-th level RNN" shape=box]
+
+  subgraph cluster0 {
+    label = "time step 0"
+
+    sent0 [label="sentence"]
+    sent1 [label="sentence"]
+
+    rnn1 [label="2-th level RNN" shape=box]
+
+    sent0 -> rnn1
+    sent1 -> rnn1
+  }
+
+  subgraph cluster1 {
+    label = "time step 1"
+
+    sent2 [label="sentence"]
+    sent3 [label="sentence"]
+
+    rnn2 [label="2-th level RNN" shape=box]
+
+    sent2 -> rnn2
+    sent3 -> rnn2
+  }
+
+  subgraph cluster2 {
+    label = "time step 2"
+
+    sent4 [label="sentence"]
+    sent5 [label="sentence"]
+
+    rnn3 [label="2-th level RNN" shape=box]
+
+    sent4 -> rnn3
+    sent5 -> rnn3
+  }
+
+
+  para0 [label="paragraph info 0"]
+  para1 [label="paragraph info 1"]
+  para2 [label="paragraph info 2"]
+
+  rnn1 -> para0
+  rnn2 -> para1
+  rnn3 -> para2
+
+  para0 -> rnn
+  para1 -> rnn
+  para2 -> rnn
+
+  chapter [label="chapter info"]
+  rnn -> chapter
+}
diff --git a/doc/design/ops/images/2_level_rnn.png b/doc/design/ops/images/2_level_rnn.png
new file mode 100644
index 0000000000..0537a75beb
Binary files /dev/null and b/doc/design/ops/images/2_level_rnn.png differ
diff --git a/doc/design/ops/images/rnn.dot b/doc/design/ops/images/rnn.dot
new file mode 100644
index 0000000000..c1141cd9c9
--- /dev/null
+++ b/doc/design/ops/images/rnn.dot
@@ -0,0 +1,87 @@
+digraph G {
+  label = "simple RNN implementation" 
+
+  ranksep=2;
+
+  //graph [nodesep=1, ranksep=1];
+
+  node[nodesep=1]
+
+  subgraph cluster0 {
+    label = "global scope"
+    rankdir = TB
+    W
+    boot_memory
+    input
+    output
+  }
+
+  subgraph cluster1 {
+    label = "step-scope 0"
+    rankdir = TB
+    memory0[label="memory"]
+    prememory0[label="pre-memory"]
+    step_input0[label="step input"]
+    step_output0[label="step output"]
+  }
+
+  subgraph cluster2 {
+    label = "step-scope 1"
+    rankdir = TB
+    memory1[label="memory"]
+    prememory1[label="pre-memory"]
+    step_input1[label="step input"]
+    step_output1[label="step output"]
+  }
+
+  subgraph cluster3 {
+    label = "step-scope 2"
+    rankdir = TB
+    memory2[label="memory"]
+    prememory2[label="pre-memory"]
+    step_input2[label="step input"]
+    step_output2[label="step output"]
+  }
+
+  stepnet [shape=box]
+  stepnet0 [shape=box, style=dashed]
+  stepnet1 [shape=box, style=dashed]
+  stepnet2 [shape=box, style=dashed]
+
+
+  edge[color=blue]
+  boot_memory -> prememory0 [label="init" color="blue"]
+  memory0 -> prememory1  [label="copy/reference" color="blue"]
+  memory1 -> prememory2 [label="copy/reference" color="blue"]
+
+  edge[color=black]
+  W -> stepnet0[constraint=false, style=dashed]
+  W -> stepnet1[constraint=false, style=dashed]
+  W -> stepnet2[constraint=false, style=dashed]
+
+  memory0 -> stepnet0[style=dashed]
+  prememory0 -> stepnet0 -> step_output0[style=dashed]
+
+  memory1 -> stepnet1[style=dashed]
+  prememory1 -> stepnet1 -> step_output1[style=dashed]
+
+  memory2 -> stepnet2[style=dashed]
+  prememory2 -> stepnet2 -> step_output2[style=dashed]
+
+  input -> step_input0
+  input -> step_input1
+  input -> step_input2
+
+  step_input0 -> stepnet0 [style=dashed]
+  step_input1 -> stepnet1[style=dashed]
+  step_input2 -> stepnet2[style=dashed]
+
+  step_output0 -> output
+  step_output1 -> output
+  step_output2 -> output
+
+  stepnet0 -> stepnet[style=dashed]
+  stepnet1 -> stepnet[style=dashed]
+  stepnet2 -> stepnet[style=dashed]
+
+}
diff --git a/doc/design/ops/images/rnn.jpg b/doc/design/ops/images/rnn.jpg
new file mode 100644
index 0000000000..9867e404cf
Binary files /dev/null and b/doc/design/ops/images/rnn.jpg differ
diff --git a/doc/design/ops/images/rnn.png b/doc/design/ops/images/rnn.png
new file mode 100644
index 0000000000..e139e373fe
Binary files /dev/null and b/doc/design/ops/images/rnn.png differ
diff --git a/doc/design/ops/images/rnn_2level_data.dot b/doc/design/ops/images/rnn_2level_data.dot
new file mode 100644
index 0000000000..1d85ae2617
--- /dev/null
+++ b/doc/design/ops/images/rnn_2level_data.dot
@@ -0,0 +1,75 @@
+digraph G {
+  chapter [label="chapter"]
+
+  subgraph cluster0 {
+    label = "paragraph 0"
+
+    top_rnn0[label="top rnn step 0" shape=box]
+
+    p0 [label="paragraph 0"]
+    p1 [label="paragraph 1"]
+  }
+
+  subgraph cluster1{
+    label = "paragraph 1"
+
+    top_rnn1[label="top rnn step 1" shape=box]
+
+    p2 [label="paragraph 0"]
+    p3 [label="paragraph 1"]
+  }
+
+  subgraph cluster_p0 {
+    label = "sentence 0"
+
+    low_rnn0 [label="low rnn step 0" shape=box]
+    s00 [label="sentence 0"]
+    s01 [label="sentence 1"]
+
+    low_rnn0 -> s00
+    low_rnn0 -> s01
+  }
+
+  subgraph cluster_p1 {
+    label = "sentence 1"
+    low_rnn1 [label="low rnn step 1" shape=box]
+    s10 [label="sentence 0"]
+    s11 [label="sentence 1"]
+    low_rnn1 -> s10
+    low_rnn1 -> s11
+  }
+
+  subgraph cluster_p2 {
+    label = "sentence 1"
+    low_rnn2 [label="low rnn step 0" shape=box]
+    s20 [label="sentence 0"]
+    s21 [label="sentence 1"]
+    low_rnn2 -> s20
+    low_rnn2 -> s21
+  }
+
+  subgraph cluster_p3 {
+    label = "sentence 1"
+    low_rnn3 [label="low rnn step 1" shape=box]
+    s30 [label="sentence 0"]
+    s31 [label="sentence 1"]
+    low_rnn3 -> s30
+    low_rnn3 -> s31
+  }
+
+
+  chapter -> top_rnn0
+  chapter -> top_rnn1
+
+  top_rnn0 -> p0
+  top_rnn0 -> p1
+  top_rnn1 -> p2
+  top_rnn1 -> p3
+
+
+  p0 -> low_rnn0
+  p1 -> low_rnn1
+  p2 -> low_rnn2
+  p3 -> low_rnn3
+
+}
diff --git a/doc/design/ops/images/rnn_2level_data.png b/doc/design/ops/images/rnn_2level_data.png
new file mode 100644
index 0000000000..4be81b2430
Binary files /dev/null and b/doc/design/ops/images/rnn_2level_data.png differ
diff --git a/doc/design/ops/rnn.md b/doc/design/ops/rnn.md
new file mode 100644
index 0000000000..a78eea7d45
--- /dev/null
+++ b/doc/design/ops/rnn.md
@@ -0,0 +1,153 @@
+# RNNOp design
+
+This document is about an RNN operator which requires that instances in a mini-batch have the same length.  We will have a more flexible RNN operator.
+
+## RNN Algorithm Implementation
+
+<p aligh="center">
+<img src="./images/rnn.jpg"/>
+</p>
+
+The above diagram shows an RNN unrolled into a full network.
+
+There are several important concepts:
+
+- *step-net*: the sub-graph to run at each step,
+- *memory*, $h_t$, the state of the current step,
+- *ex-memory*, $h_{t-1}$, the state of the previous step,
+- *initial memory value*, the ex-memory of the first step.
+
+### Step-scope
+
+There could be local variables defined in step-nets.  PaddlePaddle runtime realizes these variables in *step-scopes* -- scopes created for each step.
+
+<p aligh="center">
+<img src="./images/rnn.png"/><br/>
+Figure 2 the RNN's data flow
+</p>
+
+Please be aware that all steps run the same step-net.  Each step
+
+1. creates the step-scope,
+2. realizes local variables, including step-outputs, in the step-scope, and
+3. runs the step-net, which could use these variables.
+
+The RNN operator will compose its output from step outputs in step scopes.
+
+### Memory and Ex-memory
+
+Let's give more details about memory and ex-memory via a simply example:
+
+$$
+h_t = U h_{t-1} + W x_t
+$$,
+
+where $h_t$ and $h_{t-1}$ are the memory and ex-memory of step $t$'s respectively.
+
+In the implementation, we can make an ex-memory variable either "refers to" the memory variable of the previous step,
+or copy the value of the previous memory value to the current ex-memory variable.
+
+### Usage in Python
+
+For more information on Block, please refer to the [design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/block.md).
+
+We can define an RNN's step-net using Block:
+
+```python
+import paddle as pd
+
+X = some_op() # x is some operator's output, and is a LoDTensor
+a = some_op()
+
+# declare parameters
+W = pd.Variable(shape=[20, 30])
+U = pd.Variable(shape=[20, 30])
+
+rnn = pd.create_rnn_op(output_num=1)
+with rnn.stepnet():
+    x = rnn.add_input(X)
+    # declare a memory (rnn's step)
+    h = rnn.add_memory(init=a)
+    # h.pre_state() means previous memory of rnn
+    new_state = pd.add_two( pd.matmul(W, x) + pd.matmul(U, h.pre_state()))
+    # update current memory
+    h.update(new_state)
+    # indicate that h variables in all step scopes should be merged
+    rnn.add_outputs(h)
+
+out = rnn()
+```
+
+Python API functions in above example:
+
+- `rnn.add_input` indicates the parameter is a variable that will be segmented into step-inputs.
+- `rnn.add_memory` creates a variable used as the memory.
+- `rnn.add_outputs` mark the variables that will be concatenated across steps into the RNN output.
+
+### Nested RNN and LoDTensor
+
+An RNN whose step-net includes other RNN operators is known as an *nested RNN*.
+
+For example, we could have a 2-level RNN, where the top level corresponds to paragraphs, and the lower level corresponds to sentences.
+
+The following figure illustrates the feeding of text into the lower level, one sentence each step, and the feeding of step outputs to the top level. The final top level output is about the whole text.
+
+<p aligh="center">
+<img src="./images/2_level_rnn.png"/>
+</p>
+
+```python
+import paddle as pd
+
+W = pd.Variable(shape=[20, 30])
+U = pd.Variable(shape=[20, 30])
+
+W0 = pd.Variable(shape=[20, 30])
+U0 = pd.Variable(shape=[20, 30])
+
+# a is output of some op
+a = some_op()
+
+# chapter_data is a set of 128-dim word vectors
+# the first level of LoD is sentence
+# the second level of LoD is chapter
+chapter_data = pd.Variable(shape=[None, 128], type=pd.lod_tensor, level=2)
+
+def lower_level_rnn(paragraph):
+    '''
+    x: the input
+    '''
+    rnn = pd.create_rnn_op(output_num=1)
+    with rnn.stepnet():
+        sentence = rnn.add_input(paragraph, level=0)
+        h = rnn.add_memory(shape=[20, 30])
+        h.update(
+            pd.matmul(W, sentence) + pd.matmul(U, h.pre_state()))
+        # get the last state as sentence's info
+        rnn.add_outputs(h)
+    return rnn
+
+top_level_rnn = pd.create_rnn_op(output_num=1)
+with top_level_rnn.stepnet():
+    paragraph_data = rnn.add_input(chapter_data, level=1)
+    low_rnn = lower_level_rnn(paragraph_data)
+    paragraph_out = low_rnn()
+
+    h = rnn.add_memory(init=a)
+    h.update(
+        pd.matmul(W0, paragraph_data) + pd.matmul(U0, h.pre_state()))
+    top_level_rnn.add_outputs(h)
+
+# just output the last step
+chapter_out = top_level_rnn(output_all_steps=False)
+```
+
+in above example, the construction of the `top_level_rnn` calls  `lower_level_rnn`.  The input is a LoD Tensor. The top level RNN segments input text data into paragraphs, and the lower level RNN segments each paragraph into sentences.
+
+By default, the `RNNOp` will concatenate the outputs from all the time steps,
+if the `output_all_steps` set to False, it will only output the final time step.
+
+
+<p align="center">
+<img src="images/rnn_2level_data.png"/>
+</p>
diff --git a/doc/howto/dev/new_op_cn.md b/doc/howto/dev/new_op_cn.md
index 58665e9f2b..c6570b89ae 100644
--- a/doc/howto/dev/new_op_cn.md
+++ b/doc/howto/dev/new_op_cn.md
@@ -34,7 +34,7 @@ Kernel实现       | CPU、GPU共享Kernel实现在`.h`文件中，否则，CPU
 注册Op           | Op注册实现在`.cc`文件；Kernel注册CPU实现在`.cc`文件中，GPU实现在`.cu`文件中
 
 
-实现新的op都添加至目录[paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators)下，文件命名以`*_op.h`（如有） 、 `*_op.cc` 、`*_op.cu`（如有）结尾。
+实现新的op都添加至目录[paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators)下，文件命名以`*_op.h`（如有） 、 `*_op.cc` 、`*_op.cu`（如有）结尾。**系统会根据文件名自动构建op和其对应的Python扩展。**
 
 
 下面以矩阵乘操作，即[MulOp](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc)为例来介绍如何写带Kernel的Operator。
@@ -224,45 +224,15 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
 
 ### 5. 编译
 
-- 简单**无特殊依赖**的OP无需修改CMakeList.txt文件。[paddle/operators/CMakeLists.txt](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/CMakeLists.txt) 会自动将 `paddle/operators` 目录下新增的 `*_op.cc` 文件加入编译。
-- 较为复杂、**有额外依赖** 的operator仍需要修改[paddle/operators/CMakeLists.txt](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/CMakeLists.txt)。如，`mul_op` 依赖 `math_function`，需要在`CMakeLists.txt`中添加如下内容：
+运行下面命令可以进行编译：
 
-    ```
-    op_library(mul_op SRCS mul_op.cc mul_op.cu DEPS math_function)		 +
-    ```
-
-- 运行下面命令可以进行编译：
-
-    ```
-    make mul_op
-    ```
+```
+make mul_op
+```
 
 ## 绑定Python
 
-- 绑定Python
-
-    在 [`paddle/pybind/pybind.cc
-`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/pybind.cc) 使用`USE_OP`告知编译器需要链接的Op，具体解释参考[代码注释](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_registry.h#L81)。
-
-    ```
-    USE_OP(mul);
-    ```
-    如果只实现了CPU版本，则使用`USE_CPU_ONLY_OP`:
-
-    ```
-    USE_CPU_ONLY_OP(gather);
-    ```
-
-    如果OP不带Kernel，则使用`USE_NO_KENREL_OP`:
-
-    ```
-    USE_NO_KENREL_OP(recurrent);
-    ```
-
-
- - 生成库
-
-   无需修改 [`paddle/pybind/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/CMakeLists.txt)文件，`paddle/operators` 目录下新增的 `*_op.cc` 文件会被自动添加链接到生成的lib库中。
+系统会对新增的op自动绑定Python，并链接到生成的lib库中。
 
 ## 实现单元测试
 
@@ -354,11 +324,7 @@ class TestMulGradOp(GradientChecker):
 
 ### 编译和执行单元测试
 
-单元测试编写完成之后，在[`python/paddle/v2/framework/tests/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/CMakeLists.txt)中添加以下内容，将单元测试加入工程：
-
-```
-py_test(test_mul_op SRCS test_mul_op.py)
-```
+`python/paddle/v2/framework/tests` 目录下新增的 `test_*.py` 单元测试会被自动加入工程进行编译。
 
 请注意，**不同于Op的编译测试，运行单元测试测时需要编译整个工程**，并且编译时需要打开`WITH_TESTING`, 即`cmake paddle_dir -DWITH_TESTING=ON`。编译成功后，执行下面的命令来运行单元测试：
 
@@ -371,3 +337,10 @@ make test ARGS="-R test_mul_op -V"
 ```bash
 ctest -R test_mul_op
 ```
+
+## 注意事项
+
+- 为每个Op创建单独的`*_op.h`（如有）、`*_op.cc`和`*_op.cu`（如有）。不允许一个文件中包含多个Op，这将会导致编译出错。
+- 注册Op时的类型名，需要和该Op的名字一样。即不允许在`A_op.cc`里面，注册`REGISTER_OP(B, ...)`等，这将会导致单元测试出错。
+- 如果Op没有实现GPU Kernel，请不要创建空的`*_op.cu`，这将会导致单元测试出错。
+- 如果多个Op依赖一些共用的函数，可以创建非`*_op.*`格式的文件来存放，如`gather.h`文件。
diff --git a/doc/howto/dev/write_docs_cn.rst b/doc/howto/dev/write_docs_cn.rst
index 36e5d420c9..731a63f945 100644
--- a/doc/howto/dev/write_docs_cn.rst
+++ b/doc/howto/dev/write_docs_cn.rst
@@ -5,15 +5,13 @@
 PaddlePaddle的文档包括英文文档 ``doc`` 和中文文档 ``doc_cn`` 两个部分。文档都是通过 `cmake`_ 驱动 `sphinx`_ 编译生成，生成后的文档分别存储在编译目录的 ``doc`` 和 ``doc_cn`` 两个子目录下。
 
 
-如何构建PaddlePaddle的文档
-==========================
+如何构建文档
+============
 
-PaddlePaddle的文档构建有直接构建和基于Docker构建两种方式，我们提供了一个构建脚本build_docs.sh来进行构建。
-PaddlePaddle文档需要准备的环境相对较复杂，所以我们推荐使用基于Docker来构建PaddlePaddle的文档。
+PaddlePaddle的文档构建有两种方式。
 
-
-使用Docker构建PaddlePaddle的文档
---------------------------------
+使用Docker构建
+--------------
 
 使用Docker构建PaddlePaddle的文档，需要在系统里先安装好Docker工具包。Docker安装请参考 `Docker的官网 <https://docs.docker.com/>`_ 。安装好Docker之后可以使用源码目录下的脚本构建文档，即
 
@@ -21,58 +19,46 @@ PaddlePaddle文档需要准备的环境相对较复杂，所以我们推荐使
 
     cd TO_YOUR_PADDLE_CLONE_PATH
     cd paddle/scripts/tools/build_docs
-    bash build_docs.sh with_docker
-
-编译完成后，会在当前目录生成两个子目录\:
-
-* doc 英文文档目录
-* doc_cn 中文文档目录
+    sh build_docs.sh
 
+编译完成之后，会在当前目录生成两个子目录\: doc(英文文档目录)和 doc_cn(中文文档目录)。
 打开浏览器访问对应目录下的index.html即可访问本地文档。
 
-
-
-直接构建PaddlePaddle的文档
---------------------------
-
-因为PaddlePaddle的v2 api文档生成过程依赖于py_paddle Python包，用户需要首先确认py_paddle包已经安装。
-
-..  code-block:: bash
-
-    python -c "import py_paddle"
-
-如果提示错误，那么用户需要在本地编译安装PaddlePaddle，请参考 `源码编译文档 <http://doc.paddlepaddle.org/develop/doc/getstarted/build_and_install/build_from_source_en.html>`_ 。
-注意，用户在首次编译安装PaddlePaddle时，请将WITH_DOC选项关闭。在编译安装正确之后，请再次确认py_paddle包已经安装，即可进行下一步操作。
+直接构建
+--------
 
 如果提示正确，可以执行以下命令编译生成文档，即
 
 ..  code-block:: bash
 
     cd TO_YOUR_PADDLE_CLONE_PATH
-    cd paddle/scripts/tools/build_docs
-    bash build_docs.sh local
-
-编译完成之后，会在当前目录生成两个子目录\:
-
-* doc 英文文档目录
-* doc_cn 中文文档目录
+    mkdir -p build
+    cd build
+    cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_MKLDNN=OFF -DWITH_MKLML=OFF -DWITH_DOC=ON
+    make gen_proto_py
+    make paddle_docs paddle_docs_cn
 
+编译完成之后，会在当前目录生成两个子目录\: doc(英文文档目录)和 doc_cn(中文文档目录)。
 打开浏览器访问对应目录下的index.html即可访问本地文档。
 
 
-如何书写PaddlePaddle的文档
-==========================
+如何书写文档
+============
 
 PaddlePaddle文档使用 `sphinx`_ 自动生成，用户可以参考sphinx教程进行书写。
 
-如何更新www.paddlepaddle.org文档
-================================
+如何更新文档主题
+================
+
+PaddlePaddle文档主题在 `TO_YOUR_PADDLE_CLONE_PATH/doc_theme` 文件夹下，包含所有和前端网页设计相关的文件。
 
-开发者给PaddlePaddle代码增加的注释以PR的形式提交到github中，提交方式可参见 `贡献文档 <http://doc.paddlepaddle.org/develop/doc_cn/howto/dev/contribute_to_paddle_cn.html>`_ 。
+如何更新doc.paddlepaddle.org
+============================
+
+更新的文档以PR的形式提交到github中，提交方式参见 `贡献文档 <http://doc.paddlepaddle.org/develop/doc_cn/howto/dev/contribute_to_paddle_cn.html>`_ 。
 目前PaddlePaddle的develop分支的文档是自动触发更新的，用户可以分别查看最新的 `中文文档 <http://doc.paddlepaddle.org/develop/doc_cn/>`_ 和
 `英文文档 <http://doc.paddlepaddle.org/develop/doc/>`_ 。
 
 
-
 ..  _cmake: https://cmake.org/
 ..  _sphinx: http://www.sphinx-doc.org/en/1.4.8/
diff --git a/paddle/capi/CMakeLists.txt b/paddle/capi/CMakeLists.txt
index dde99ab340..3af111eb57 100644
--- a/paddle/capi/CMakeLists.txt
+++ b/paddle/capi/CMakeLists.txt
@@ -64,9 +64,29 @@ link_paddle_exe(paddle_capi_shared)
 install(FILES ${CAPI_HEADERS} DESTINATION include/paddle)
 install(FILES ${CMAKE_CURRENT_BINARY_DIR}/config.h DESTINATION include/paddle)
 if(ANDROID)
+  execute_process(
+    COMMAND ${GIT_EXECUTABLE} log --pretty=oneline -1
+    OUTPUT_VARIABLE GIT_COMMITS_LIST
+    RESULT_VARIABLE GIT_COMMITS_LIST_RESULT
+    ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
+  if(${GIT_COMMITS_LIST_RESULT})
+    set(GIT_COMMITS_LIST "No commits.")
+  endif()
   install(FILES ${CMAKE_CURRENT_BINARY_DIR}/${capi_whole_library}
           DESTINATION lib/${ANDROID_ABI})
   install(TARGETS paddle_capi_shared DESTINATION lib/${ANDROID_ABI})
+  install(CODE "FILE(WRITE ${CMAKE_INSTALL_PREFIX}/lib/${ANDROID_ABI}/BUILD.txt
+          \"Compiler:\n\"
+          \"\\t${CMAKE_C_COMPILER}\\n\"
+          \"\\t${CMAKE_CXX_COMPILER}\\n\"
+          \"Compiler Flags:\\n\"
+          \"\\t${CMAKE_F_FLAGS}\\n\"
+          \"\\t${CMAKE_CXX_FLAGS}\\n\"
+          \"Android API: ${CMAKE_SYSTEM_VERSION}\\n\"
+          \"Lastest commit:\\n\"
+          \"\\t${GIT_COMMITS_LIST}\\n\"
+      )"
+  )
 else(ANDROID)
   install(FILES ${CMAKE_CURRENT_BINARY_DIR}/${capi_whole_library} DESTINATION lib)
   install(TARGETS paddle_capi_shared DESTINATION lib)
diff --git a/paddle/cuda/include/hl_cuda_cudnn.h b/paddle/cuda/include/hl_cuda_cudnn.h
index 3f68c62de6..b44b071bd1 100644
--- a/paddle/cuda/include/hl_cuda_cudnn.h
+++ b/paddle/cuda/include/hl_cuda_cudnn.h
@@ -22,10 +22,10 @@ limitations under the License. */
  */
 typedef enum {
   HL_POOLING_MAX = 0,
-  // average includes padded values
-  HL_POOLING_AVERAGE = 1,
   // average does not include padded values
-  HL_POOLING_AVERAGE_EXCLUDE_PADDING = 2,
+  HL_POOLING_AVERAGE = 1,
+  // average includes padded values
+  HL_POOLING_AVERAGE_INCLUDE_PADDING = 2,
   HL_POOLING_END
 } hl_pooling_mode_t;
 
diff --git a/paddle/cuda/include/hl_tensor_ops.h b/paddle/cuda/include/hl_tensor_ops.h
index 93d38b7d22..b2bf334dab 100644
--- a/paddle/cuda/include/hl_tensor_ops.h
+++ b/paddle/cuda/include/hl_tensor_ops.h
@@ -461,7 +461,7 @@ class add<float32x4_t> {
 public:
   INLINE float32x4_t operator()(const float32x4_t a,
                                 const float32x4_t b) const {
-    return vmulq_f32(a, b);
+    return vaddq_f32(a, b);
   }
 };
 
diff --git a/paddle/cuda/src/hl_cuda_cnn.cu b/paddle/cuda/src/hl_cuda_cnn.cu
index 9ba3d14261..58674febdc 100644
--- a/paddle/cuda/src/hl_cuda_cnn.cu
+++ b/paddle/cuda/src/hl_cuda_cnn.cu
@@ -211,13 +211,11 @@ __global__ void KeAvgPoolForward(const int nthreads,
 
     int hstart = ph * strideH - padH;
     int wstart = pw * strideW - padW;
-    int hend = min(hstart + sizeY, height + padH);
-    int wend = min(wstart + sizeX, width + padW);
-    int pool_size = (hend - hstart) * (wend - wstart);
+    int hend = min(hstart + sizeY, height);
+    int wend = min(wstart + sizeX, width);
     hstart = max(hstart, 0);
     wstart = max(wstart, 0);
-    hend = min(hend, height);
-    wend = min(wend, width);
+    int pool_size = (hend - hstart) * (wend - wstart);
 
     real aveval = 0;
     inputData += (frameNum * channels + c) * height * width;
@@ -299,12 +297,14 @@ __global__ void KeAvgPoolBackward(const int nthreads,
     outGrad += (frameNum * outStride + offsetC * pooledH * pooledW);
 
     for (int ph = phstart; ph < phend; ++ph) {
+      int hstart = ph * strideH - padH;
+      int hend = min(hstart + sizeY, height);
+      hstart = max(hstart, 0);
       for (int pw = pwstart; pw < pwend; ++pw) {
         // figure out the pooling size
-        int hstart = ph * strideH - padH;
         int wstart = pw * strideW - padW;
-        int hend = min(hstart + sizeY, height + padH);
-        int wend = min(wstart + sizeX, width + padW);
+        int wend = min(wstart + sizeX, width);
+        wstart = max(wstart, 0);
         int poolsize = (hend - hstart) * (wend - wstart);
         gradient += outGrad[ph * pooledW + pw] / poolsize;
       }
@@ -600,16 +600,13 @@ __global__ void KeAvgPool3DForward(const int nthreads,
     int dstart = pd * strideD - padD;
     int hstart = ph * strideH - padH;
     int wstart = pw * strideW - padW;
-    int dend = min(dstart + sizeZ, depth + padD);
-    int hend = min(hstart + sizeY, height + padH);
-    int wend = min(wstart + sizeX, width + padW);
-    int pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart);
+    int dend = min(dstart + sizeZ, depth);
+    int hend = min(hstart + sizeY, height);
+    int wend = min(wstart + sizeX, width);
     dstart = max(dstart, 0);
     hstart = max(hstart, 0);
     wstart = max(wstart, 0);
-    dend = min(dend, depth);
-    hend = min(hend, height);
-    wend = min(wend, width);
+    int pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart);
 
     real aveval = 0;
     inputData += (frameNum * channels + c) * depth * height * width;
@@ -712,15 +709,18 @@ __global__ void KeAvgPool3DBackward(const int nthreads,
     outGrad += (frameNum * channels + offsetC) * pooledD * pooledH * pooledW;
 
     for (int pd = pdstart; pd < pdend; ++pd) {
+      int dstart = pd * strideD - padD;
+      int dend = min(dstart + sizeZ, depth);
+      dstart = max(dstart, 0);
       for (int ph = phstart; ph < phend; ++ph) {
+        int hstart = ph * strideH - padH;
+        int hend = min(hstart + sizeY, height);
+        hstart = max(hstart, 0);
         for (int pw = pwstart; pw < pwend; ++pw) {
           // figure out the pooling size
-          int dstart = pd * strideD - padD;
-          int hstart = ph * strideH - padH;
           int wstart = pw * strideW - padW;
-          int dend = min(dstart + sizeZ, depth + padD);
-          int hend = min(hstart + sizeY, height + padH);
-          int wend = min(wstart + sizeX, width + padW);
+          int wend = min(wstart + sizeX, width);
+          wstart = max(wstart, 0);
           int poolsize = (dend - dstart) * (hend - hstart) * (wend - wstart);
           gradient += outGrad[(pd * pooledH + ph) * pooledW + pw] / poolsize;
         }
diff --git a/paddle/cuda/src/hl_cuda_cudnn.cc b/paddle/cuda/src/hl_cuda_cudnn.cc
index f38ef69255..b8caf48f9c 100644
--- a/paddle/cuda/src/hl_cuda_cudnn.cc
+++ b/paddle/cuda/src/hl_cuda_cudnn.cc
@@ -432,11 +432,11 @@ void hl_create_pooling_descriptor(hl_pooling_descriptor* pooling_desc,
       cudnn_mode = CUDNN_POOLING_MAX;
       break;
     case HL_POOLING_AVERAGE:
-      cudnn_mode = CUDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING;
-      break;
-    case HL_POOLING_AVERAGE_EXCLUDE_PADDING:
       cudnn_mode = CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING;
       break;
+    case HL_POOLING_AVERAGE_INCLUDE_PADDING:
+      cudnn_mode = CUDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING;
+      break;
     default:
       LOG(FATAL) << "parameter mode error";
   }
diff --git a/paddle/framework/CMakeLists.txt b/paddle/framework/CMakeLists.txt
index c0838d9b75..3371962c63 100644
--- a/paddle/framework/CMakeLists.txt
+++ b/paddle/framework/CMakeLists.txt
@@ -9,6 +9,7 @@ cc_test(eigen_test SRCS eigen_test.cc DEPS tensor)
 
 cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor)
 cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor)
+nv_test(lod_tensor_gpu_test SRCS lod_tensor_test.cu DEPS lod_tensor)
 
 cc_test(variable_test SRCS variable_test.cc)
 
diff --git a/paddle/framework/backward.md b/paddle/framework/backward.md
index c762811dfc..0a6d762bc8 100644
--- a/paddle/framework/backward.md
+++ b/paddle/framework/backward.md
@@ -2,11 +2,22 @@
 
 ## Motivation
 
-In Neural Network, the backpropagation algorithm follows the chain rule, so we need to compound the gradient operators/expressions together with the chain rule. Every forward network needs a backward network to construct the full computation graph, the operator/expression's backward pass will be generated respect to forward pass.
+In Neural Network, many model is solved by the the backpropagation algorithm(known as BP) at present. Technically it caculates the gradient of the loss function, then distributed back through the networks. Follows the chain rule, so we need a module chains the gradient operators/expressions together with to construct the backward pass. Every forward network needs a backward network to construct the full computation graph, the operator/expression's backward pass will be generated respect to forward pass. 
 
-## Backward Operator Registry
+## Implementation
 
-A backward network is built up with several backward operators. Backward operators take forward operators' inputs outputs, and output gradients and then calculate its input gradients.
+In this design doc, we exported only one API for generating the backward pass.
+
+```c++
+std::unique_ptr<OperatorBase> Backward(const OperatorBase& forwardOp,
+    const std::unordered_set<std::string>& no_grad_vars);
+```
+
+The implementation behind it can be divided into two parts, **Backward Operator Creating** and **Backward Operator Building**.
+
+### Backward Operator Registry
+
+A backward network is built up with several backward operators. Backward operators take forward operators' inputs, outputs, and output gradients and then calculate its input gradients.
 
 |                        | forward operator | backward operator 
 | ---------------------- | ---------------- |------------------------- |		
@@ -25,7 +36,7 @@ REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad);
 
 `mul_grad` is the type of backward operator, and `MulOpGrad` is its class name.
 
-## Backward Opeartor Creating
+### Backward Opeartor Creating
 
 Given a certain forward operator, we can get its corresponding backward operator by calling:
 
@@ -43,40 +54,47 @@ The function `BuildGradOp` will sequentially execute following processes:
 
 4. Building backward operator with `inputs`, `outputs` and forward operator's attributes.
 
-## Backward Network Building
-
-A backward network is a series of backward operators. The main idea of building a backward network is creating backward operators in the inverted sequence and put them together.
+### Backward Network Building
 
-In our design, the network itself is also a kind of operator. So the operators contained by a big network may be some small network. 
-
-given a forward network, it generates the backward network. We only care about the Gradients—`OutputGradients`, `InputGradients`.
+A backward network is a series of backward operators. The main idea of building a backward network is creating backward operators in the inverted sequence and append them together one by one. There is some corner case need to process specially.
 
 1. Op 
 
-   when the input forward network is an Op, return its gradient Operator Immediately.
+   When the input forward network is an Op, return its gradient Operator Immediately. If all of its outputs are in no gradient set, then return a special `NOP`.
 
 2. NetOp 
 
-   when the input forward network is a NetOp, it needs to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to the forward NetOp.
+   In our design, the network itself is also a kind of operator(**NetOp**). So the operators contained by a big network may be some small network. When the input forward network is a NetOp, it needs to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to the forward NetOp.
+
+3. RnnOp
+
+   RnnOp is a nested stepnet operator.  Backward module need to recusively call `Backward` for every stepnet.
+
+4. Sharing Variables
+
+   **sharing variables**. As illustrated in the pictures, two operator's share the same variable name of W@GRAD, which will overwrite their sharing input variable. 
+
+<p align="center">
+<img src="./images/duplicate_op.png" width="50%" ><br/>
 
-   **shared variable**. As illustrated in the pictures, two operator's `Output` `Gradient` will overwrite their shared input variable.  
+​	pic 1. Sharing variables in operators. 
 
-   <p align="center">
-   <img src="./images/duplicate_op.png" width="50%" ><br/>
+</p>
 
-   1. Shared variable in operators. 
+​	Sharing variable between operators or same input variable used in multiple operators leads to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively and add a generic add operator to replace the overwrite links. 
 
-   </p>
+<p align="center">
+<img src="images/duplicate_op2.png" width="40%" ><br/>
 
-   Share variable between operators or same input variable used in multiple operators leads to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively and add a generic add operator replace the overwrite links. 
+​	pic 2. Replace sharing variable's gradient with `Add` operator.
 
-   <p align="center">
-   <img src="images/duplicate_op2.png" width="50%" ><br/>
+</p>
 
-   2. Replace shared variable's gradient with `Add` operator.
+​	Because our framework finds variables accord to their names, we need to rename the output links. We add a suffix of number to represent its position in clockwise. 
 
-   </p>
+5. Part of Gradient is Zero.
 
+   In the whole graph, there is some case of that one operator's gradient is not needed, but its input's gradient is a dependency link of other operator,  we need to fill a same shape gradient matrix in the position. In our implement, we insert a special `fillZeroLike` operator.
 
 
-​	Then collect the sub graph `OutputGradients`/`InputGradients` as the NetOp's and return it.
+Follow these rules above, then collect the sub graph `OutputGradients`/`InputGradients` as the NetOp's and return it.
diff --git a/paddle/framework/images/duplicate_op2.graffle b/paddle/framework/images/duplicate_op2.graffle
index ede3bca30a..5cec3bc64d 100644
Binary files a/paddle/framework/images/duplicate_op2.graffle and b/paddle/framework/images/duplicate_op2.graffle differ
diff --git a/paddle/framework/images/duplicate_op2.png b/paddle/framework/images/duplicate_op2.png
index 4e872dc2ca..21cdd5cabf 100644
Binary files a/paddle/framework/images/duplicate_op2.png and b/paddle/framework/images/duplicate_op2.png differ
diff --git a/paddle/framework/lod_tensor.h b/paddle/framework/lod_tensor.h
index 154068fef6..fac5cd20aa 100644
--- a/paddle/framework/lod_tensor.h
+++ b/paddle/framework/lod_tensor.h
@@ -18,8 +18,10 @@
 #ifndef PADDLE_ONLY_CPU
 #include <thrust/device_vector.h>
 #include <thrust/host_vector.h>
+#include <thrust/system/cuda/experimental/pinned_allocator.h>
 #endif
 
+#include <glog/logging.h>
 #include "paddle/framework/ddim.h"
 #include "paddle/framework/tensor.h"
 #include "paddle/platform/enforce.h"
@@ -32,7 +34,8 @@ template <typename T>
 using Vector = std::vector<T>;
 #else
 template <typename T>
-using Vector = thrust::host_vector<T>;
+using Vector = thrust::host_vector<
+    T, thrust::system::cuda::experimental::pinned_allocator<T>>;
 #endif
 
 using LoD = std::vector<Vector<size_t>>;
@@ -48,18 +51,15 @@ bool operator==(const LoD& a, const LoD& b);
  * LoDTensor (Level of details Tensor)
  * see https://en.wikipedia.org/wiki/Level_of_details for reference.
  */
-class LoDTensor {
+class LoDTensor : public Tensor {
  public:
   LoDTensor() {}
-  LoDTensor(const LoD& lod, Tensor* t) : lod_(lod), tensor_(t) {}
 
-  void set_lod(const LoD& lod) { lod_ = lod; }
-
-  void set_tensor(Tensor* tensor) { tensor_ = tensor; }
+  explicit LoDTensor(const LoD& lod) : lod_(lod) {}
 
-  Tensor& tensor() { return *tensor_; }
+  void set_lod(const LoD& lod) { lod_ = lod; }
 
-  LoD lod() { return lod_; }
+  LoD lod() const { return lod_; }
 
   /*
    * Get a element from LoD.
@@ -101,7 +101,6 @@ class LoDTensor {
 
  private:
   LoD lod_;
-  Tensor* tensor_;  // not owned
 };
 }  // namespace framework
 }  // namespace paddle
diff --git a/paddle/framework/lod_tensor_test.cc b/paddle/framework/lod_tensor_test.cc
index 1da8553134..7915326b27 100644
--- a/paddle/framework/lod_tensor_test.cc
+++ b/paddle/framework/lod_tensor_test.cc
@@ -36,69 +36,64 @@ class LoDTensorTester : public ::testing::Test {
 
     ASSERT_EQ(lod.size(), 3UL);
 
-    tensor.Resize({20 /*batch size*/, 128 /*dim*/});
+    lod_tensor_.Resize({20 /*batch size*/, 128 /*dim*/});
     // malloc memory
-    tensor.mutable_data<float>(place);
+    lod_tensor_.mutable_data<float>(place);
 
-    lod_tensor.set_lod(lod);
-    lod_tensor.set_tensor(&tensor);
+    lod_tensor_.set_lod(lod);
   }
 
  protected:
   platform::CPUPlace place;
-  Tensor tensor;
-  LoDTensor lod_tensor;
+  LoDTensor lod_tensor_;
 };
 
-TEST_F(LoDTensorTester, NumLevels) { ASSERT_EQ(lod_tensor.NumLevels(), 3UL); }
+TEST_F(LoDTensorTester, NumLevels) { ASSERT_EQ(lod_tensor_.NumLevels(), 3UL); }
 
 TEST_F(LoDTensorTester, NumElements) {
-  ASSERT_EQ(lod_tensor.NumElements(0), 2UL);
-  ASSERT_EQ(lod_tensor.NumElements(1), 4UL);
-  ASSERT_EQ(lod_tensor.NumElements(2), 8UL);
+  ASSERT_EQ(lod_tensor_.NumElements(0), 2UL);
+  ASSERT_EQ(lod_tensor_.NumElements(1), 4UL);
+  ASSERT_EQ(lod_tensor_.NumElements(2), 8UL);
 }
 
 TEST_F(LoDTensorTester, SliceLevels) {
   // slice 1 level
   for (size_t level = 0; level < 3UL; ++level) {
-    LoDTensor new_lod_tensor = lod_tensor;
+    LoDTensor new_lod_tensor = lod_tensor_;
     new_lod_tensor.SliceLevels(level, level + 1);
     ASSERT_EQ(new_lod_tensor.NumLevels(), 1UL);
-    ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor.NumElements(level));
-    ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
-              lod_tensor.tensor().data<float>());
+    ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor_.NumElements(level));
+    ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
   }
   // slice 2 level
   for (size_t level = 0; level < 2UL; ++level) {
-    LoDTensor new_lod_tensor = lod_tensor;
+    LoDTensor new_lod_tensor = lod_tensor_;
     new_lod_tensor.SliceLevels(level, level + 2);
     ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
-    ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor.NumElements(level));
-    ASSERT_EQ(new_lod_tensor.NumElements(1), lod_tensor.NumElements(level + 1));
-    ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
-              lod_tensor.tensor().data<float>());
+    ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor_.NumElements(level));
+    ASSERT_EQ(new_lod_tensor.NumElements(1),
+              lod_tensor_.NumElements(level + 1));
+    ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
   }
 }
 
 TEST_F(LoDTensorTester, SliceInLevel) {
   size_t level = 0;
-  LoDTensor new_lod_tensor = lod_tensor;
+  LoDTensor new_lod_tensor = lod_tensor_;
   new_lod_tensor.SliceInLevel(level, 0, 2);
   EXPECT_EQ(new_lod_tensor.NumLevels(), 3UL);
   EXPECT_EQ(new_lod_tensor.NumElements(0), 2UL);
   EXPECT_EQ(new_lod_tensor.NumElements(1), 4UL);
   EXPECT_EQ(new_lod_tensor.NumElements(2), 8UL);
-  ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
-            lod_tensor.tensor().data<float>());
+  ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
 
   level = 1;
-  new_lod_tensor = lod_tensor;
+  new_lod_tensor = lod_tensor_;
   new_lod_tensor.SliceInLevel(level, 0, 2);
   ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
   ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL);
   ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL);
-  ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
-            lod_tensor.tensor().data<float>());
+  ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
 }
 
 }  // namespace framework
diff --git a/paddle/framework/lod_tensor_test.cu b/paddle/framework/lod_tensor_test.cu
new file mode 100644
index 0000000000..97e69cdb2e
--- /dev/null
+++ b/paddle/framework/lod_tensor_test.cu
@@ -0,0 +1,50 @@
+/*
+  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 <cuda.h>
+#include <cuda_runtime.h>
+#include "paddle/framework/lod_tensor.h"
+#include "paddle/platform/assert.h"
+
+#include <gtest/gtest.h>
+
+__global__ void test(size_t* a, int size) {
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < size;
+       i += blockDim.x * gridDim.x) {
+    a[i] *= 2;
+  }
+}
+
+TEST(LoDTensor, LoDInGPU) {
+  paddle::framework::LoDTensor lod_tensor;
+  paddle::platform::GPUPlace place(0);
+
+  paddle::framework::LoD src_lod;
+  src_lod.push_back(std::vector<size_t>{0, 2, 4, 6, 8, 10, 12, 14});
+
+  lod_tensor.Resize({14, 16});
+  lod_tensor.mutable_data<float>(place);
+
+  lod_tensor.set_lod(src_lod);
+  CHECK_EQ(lod_tensor.lod_element(0, 2), 4);
+  CHECK_EQ(lod_tensor.lod_element(0, 4), 8);
+
+  auto lod = lod_tensor.lod();
+
+  test<<<1, 8>>>(lod[0].data(), lod[0].size());
+  cudaDeviceSynchronize();
+
+  for (size_t i = 0; i < src_lod[0].size(); ++i) {
+    CHECK_EQ(lod[0].data()[i], src_lod[0].data()[i] * 2);
+  }
+}
diff --git a/paddle/framework/operator.cc b/paddle/framework/operator.cc
index 790cfc4746..c57537be4b 100644
--- a/paddle/framework/operator.cc
+++ b/paddle/framework/operator.cc
@@ -123,6 +123,15 @@ OperatorBase::OperatorBase(const std::string& type,
   CheckAllInputOutputSet();
 }
 
+std::vector<std::string> OperatorBase::InputVars() const {
+  std::vector<std::string> ret_val;
+  for (auto& o : outputs_) {
+    ret_val.reserve(ret_val.size() + o.second.size());
+    ret_val.insert(ret_val.end(), o.second.begin(), o.second.end());
+  }
+  return ret_val;
+}
+
 std::vector<std::string> OperatorBase::OutputVars(bool has_intermediate) const {
   std::vector<std::string> ret_val;
   if (has_intermediate) {
@@ -177,6 +186,48 @@ void OperatorBase::GenerateTemporaryNames() {
   }
 }
 
+template <>
+const Tensor* InferShapeContext::Input<Tensor>(const std::string& name) const {
+  auto* var = InputVar(name);
+  return var == nullptr ? nullptr : GetTensorFromVar(var);
+}
+
+template <>
+const std::vector<const Tensor*> InferShapeContext::MultiInput<Tensor>(
+    const std::string& name) const {
+  auto names = op().Inputs(name);
+  std::vector<const Tensor*> res;
+  res.reserve(names.size());
+  std::transform(names.begin(), names.end(), std::back_inserter(res),
+                 [&](const std::string& sub_name) {
+                   auto var = scope_.FindVar(sub_name);
+                   return var == nullptr ? nullptr : GetTensorFromVar(var);
+                 });
+  return res;
+}
+
+template <>
+Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const {
+  auto* var = OutputVar(name);
+  return var == nullptr ? nullptr : const_cast<Tensor*>(GetTensorFromVar(var));
+}
+
+template <>
+std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
+    const std::string& name) const {
+  auto names = op().Outputs(name);
+  std::vector<Tensor*> res;
+  res.reserve(names.size());
+  std::transform(names.begin(), names.end(), std::back_inserter(res),
+                 [&](const std::string& sub_name) {
+                   auto var = scope().FindVar(sub_name);
+                   return var == nullptr
+                              ? nullptr
+                              : const_cast<Tensor*>(GetTensorFromVar(var));
+                 });
+  return res;
+}
+
 void OpProtoAndCheckerMaker::Validate() {
   validated_ = true;
   CheckNoDuplicatedInOutAttrs();
diff --git a/paddle/framework/operator.h b/paddle/framework/operator.h
index 9a98d4d3be..adae7bfc3d 100644
--- a/paddle/framework/operator.h
+++ b/paddle/framework/operator.h
@@ -22,6 +22,7 @@ limitations under the License. */
 #include "op_info.h"
 #include "paddle/framework/attribute.h"
 #include "paddle/framework/framework.pb.h"
+#include "paddle/framework/lod_tensor.h"
 #include "paddle/framework/scope.h"
 #include "paddle/framework/tensor.h"
 #include "paddle/platform/device_context.h"
@@ -94,11 +95,14 @@ class OperatorBase {
 
   const VariableNameMap& Inputs() const { return inputs_; }
   const VariableNameMap& Outputs() const { return outputs_; }
+
   //! Get a input with argument's name described in `op_proto`
   std::string Input(const std::string& name) const;
   //! Get a input which has multiple variables.
   const std::vector<std::string>& Inputs(const std::string& name) const;
 
+  std::vector<std::string> InputVars() const;
+
   //! Get a output with argument's name described in `op_proto`
   std::string Output(const std::string& name) const;
   //! Get an output which has multiple variables.
@@ -311,9 +315,9 @@ class InferShapeContext {
   }
 
   template <typename T>
-  std::vector<const T*> MultiOutput(const std::string& name) const {
+  std::vector<T*> MultiOutput(const std::string& name) const {
     auto names = op_.Outputs(name);
-    std::vector<const T*> res;
+    std::vector<T*> res;
     res.reserve(names.size());
     std::transform(names.begin(), names.end(), std::back_inserter(res),
                    [&](const std::string& sub_name) {
@@ -323,11 +327,27 @@ class InferShapeContext {
     return res;
   }
 
+  const Tensor* GetTensorFromVar(const Variable* var) const {
+    if (var->IsType<LoDTensor>()) {
+      return &var->Get<LoDTensor>();
+    }
+    PADDLE_ENFORCE(var->IsType<Tensor>(),
+                   "The Input(%s) must be LoDTensor or Tensor.");
+    return &var->Get<Tensor>();
+  }
+
  private:
   const OperatorBase& op_;
   const Scope& scope_;
 };
 
+template <>
+const Tensor* InferShapeContext::Input<Tensor>(const std::string& name) const;
+
+template <>
+const std::vector<const Tensor*> InferShapeContext::MultiInput<Tensor>(
+    const std::string& name) const;
+
 template <typename T>
 struct EigenDeviceConverter;
 
@@ -360,9 +380,37 @@ class ExecutionContext : public InferShapeContext {
     return device_context_;
   }
 
+  // redefine Output function,
+  // use Variable::Get instead of Variable::GetMutable
+  template <typename T>
+  T* Output(const std::string& name) const {
+    auto var = OutputVar(name);
+    return var == nullptr ? nullptr : const_cast<T*>(&var->Get<T>());
+  }
+
+  // redefine MultiOutput function.
+  // use Variable::Get instead of Variable::GetMutable
+  template <typename T>
+  std::vector<T*> MultiOutput(const std::string& name) const {
+    auto names = op().Outputs(name);
+    std::vector<T*> res;
+    res.reserve(names.size());
+    std::transform(
+        names.begin(), names.end(), std::back_inserter(res),
+        [&](const std::string& sub_name) { return Output<T>(sub_name); });
+    return res;
+  }
+
   const platform::DeviceContext* device_context_;
 };
 
+template <>
+Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const;
+
+template <>
+std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
+    const std::string& name) const;
+
 class OpKernel {
  public:
   /**
diff --git a/paddle/framework/tensor.h b/paddle/framework/tensor.h
index ce938b2143..4b5a2ae523 100644
--- a/paddle/framework/tensor.h
+++ b/paddle/framework/tensor.h
@@ -81,6 +81,9 @@ class Tensor {
   /*! Return the dimensions of the memory block. */
   inline const DDim& dims() const;
 
+  /*! Return the numel of the memory block. */
+  inline int64_t numel() const;
+
   /*! Resize the dimensions of the memory block. */
   inline Tensor& Resize(const DDim& dims);
 
@@ -162,6 +165,12 @@ class Tensor {
   /*! points to dimensions of memory block. */
   DDim dims_;
 
+  /**
+   * A cache of the number of elements in a tensor.
+   * Would be 0 for an uninitialized tensor.
+   */
+  int64_t numel_;
+
   /**
    * @brief   A PlaceHolder may be shared by more than one tensor.
    *
diff --git a/paddle/framework/tensor_impl.h b/paddle/framework/tensor_impl.h
index 637f04ae00..ed166935f7 100644
--- a/paddle/framework/tensor_impl.h
+++ b/paddle/framework/tensor_impl.h
@@ -22,9 +22,9 @@ namespace framework {
 template <typename T>
 inline void Tensor::check_memory_size() const {
   PADDLE_ENFORCE_NOT_NULL(
-      holder_, "Tenosr holds no memory. Call Tensor::mutable_data first.");
+      holder_, "Tensor holds no memory. Call Tensor::mutable_data first.");
   PADDLE_ENFORCE_GE(
-      holder_->size(), product(dims_) * sizeof(T) + offset_,
+      holder_->size(), numel() * sizeof(T) + offset_,
       "Tensor's dims_ is out of bound. Call Tensor::mutable_data "
       "first to re-allocate memory.\n"
       "or maybe the required data-type mismatches the data already stored.");
@@ -54,11 +54,11 @@ inline T* Tensor::mutable_data(DDim dims, platform::Place place) {
 template <typename T>
 inline T* Tensor::mutable_data(platform::Place place) {
   static_assert(std::is_pod<T>::value, "T must be POD");
-  PADDLE_ENFORCE_GT(product(dims_), 0,
+  PADDLE_ENFORCE_GT(numel(), 0,
                     "Tensor's numel must be larger than zero to call "
                     "Tensor::mutable_data. Call Tensor::set_dim first.");
   /* some versions of boost::variant don't have operator!= */
-  int64_t size = product(dims_) * sizeof(T);
+  int64_t size = numel() * sizeof(T);
   if (holder_ == nullptr || !(holder_->place() == place) ||
       holder_->size() < size + offset_) {
     if (platform::is_cpu_place(place)) {
@@ -97,7 +97,7 @@ inline void Tensor::CopyFrom(const Tensor& src,
 
   auto dst_ptr = static_cast<void*>(mutable_data<T>(dst_place));
 
-  auto size = product(src.dims_) * sizeof(T);
+  auto size = src.numel() * sizeof(T);
 
   if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
     memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
@@ -131,7 +131,7 @@ inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {
   PADDLE_ENFORCE_LT(begin_idx, end_idx,
                     "Begin index must be less than end index.");
   PADDLE_ENFORCE_NE(dims_[0], 1, "Can not slice a tensor with dims_[0] = 1.");
-  size_t base = product(dims_) / dims_[0];
+  size_t base = numel() / dims_[0];
   Tensor dst;
   dst.holder_ = holder_;
   DDim dst_dims = dims_;
@@ -143,11 +143,14 @@ inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {
 
 inline Tensor& Tensor::Resize(const DDim& dims) {
   dims_ = dims;
+  numel_ = product(dims_);
   return *this;
 }
 
 inline const DDim& Tensor::dims() const { return dims_; }
 
+inline int64_t Tensor::numel() const { return numel_; }
+
 template <typename T>
 inline Tensor ReshapeToMatrix(const Tensor& src, int num_col_dims) {
   Tensor res;
diff --git a/paddle/framework/tensor_test.cc b/paddle/framework/tensor_test.cc
index 55302ea471..e2ec738de3 100644
--- a/paddle/framework/tensor_test.cc
+++ b/paddle/framework/tensor_test.cc
@@ -36,7 +36,7 @@ TEST(Tensor, DataAssert) {
   } catch (paddle::platform::EnforceNotMet err) {
     caught = true;
     std::string msg =
-        "holder_ should not be null\nTenosr holds no memory. Call "
+        "holder_ should not be null\nTensor holds no memory. Call "
         "Tensor::mutable_data first.";
     const char* what = err.what();
     for (size_t i = 0; i < msg.length(); ++i) {
@@ -112,7 +112,7 @@ TEST(Tensor, ShareDataWith) {
     } catch (paddle::platform::EnforceNotMet err) {
       caught = true;
       std::string msg =
-          "holder_ should not be null\nTenosr holds no memory. Call "
+          "holder_ should not be null\nTensor holds no memory. Call "
           "Tensor::mutable_data first.";
       const char* what = err.what();
       for (size_t i = 0; i < msg.length(); ++i) {
@@ -274,4 +274,4 @@ TEST(Tensor, ReshapeToMatrix) {
   Tensor res = ReshapeToMatrix<int>(src, 2);
   ASSERT_EQ(res.dims()[0], 2 * 3);
   ASSERT_EQ(res.dims()[1], 4 * 9);
-}
\ No newline at end of file
+}
diff --git a/paddle/function/neon/NeonDepthwiseConv.h b/paddle/function/neon/NeonDepthwiseConv.h
index aefeea78ba..33722d3cac 100644
--- a/paddle/function/neon/NeonDepthwiseConv.h
+++ b/paddle/function/neon/NeonDepthwiseConv.h
@@ -594,7 +594,7 @@ struct StridePadding {
         float32x4_t s1 = vdupq_n_f32(0.f);
         for (int s = 0; s < step; s++) {
           float32x4_t s0 = vld1q_f32(input);
-          float32x4x2_t v = {s0, s1};
+          float32x4x2_t v = {{s0, s1}};
           vst2q_f32(inputPadding, v);
           input += 4;
           inputPadding += 8;
diff --git a/paddle/gserver/layers/CudnnPoolLayer.cpp b/paddle/gserver/layers/CudnnPoolLayer.cpp
index 4adb2d4709..810a1af2d0 100644
--- a/paddle/gserver/layers/CudnnPoolLayer.cpp
+++ b/paddle/gserver/layers/CudnnPoolLayer.cpp
@@ -29,9 +29,9 @@ bool CudnnPoolLayer::typeCheck(const std::string &poolType,
     if (mode) {
       *mode = HL_POOLING_AVERAGE;
     }
-  } else if (poolType == "cudnn-avg-excl-pad-pool") {
+  } else if (poolType == "cudnn-avg-incl-pad-pool") {
     if (mode) {
-      *mode = HL_POOLING_AVERAGE_EXCLUDE_PADDING;
+      *mode = HL_POOLING_AVERAGE_INCLUDE_PADDING;
     }
   } else {
     return false;
diff --git a/paddle/gserver/layers/DeConv3DLayer.cpp b/paddle/gserver/layers/DeConv3DLayer.cpp
index 1b59ed60c5..3eea638649 100644
--- a/paddle/gserver/layers/DeConv3DLayer.cpp
+++ b/paddle/gserver/layers/DeConv3DLayer.cpp
@@ -53,27 +53,27 @@ bool DeConv3DLayer::init(const LayerMap &layerMap,
 
 size_t DeConv3DLayer::getSize() {
   CHECK_NE(inputLayers_.size(), 0UL);
-  outputH_.clear();
-  outputW_.clear();
-  outputD_.clear();
+  imgSizeW_.clear();
+  imgSizeH_.clear();
+  imgSizeD_.clear();
   N_.clear();
   NOut_.clear();
   size_t layerSize = 0;
   for (size_t i = 0; i < inputLayers_.size(); ++i) {
-    outputW_.push_back(
-        imageSize(imgSizeW_[i], filterSize_[i], padding_[i], stride_[i], true));
-    outputH_.push_back(imageSize(
-        imgSizeH_[i], filterSizeY_[i], paddingY_[i], strideY_[i], true));
-    outputD_.push_back(imageSize(
-        imgSizeD_[i], filterSizeZ_[i], paddingZ_[i], strideZ_[i], true));
-    NOut_.push_back(outputD_[i] * outputH_[i] * outputW_[i]);
-    N_.push_back(imgSizeD_[i] * imgSizeH_[i] * imgSizeW_[i]);
+    imgSizeW_.push_back(
+        imageSize(outputW_[i], filterSize_[i], padding_[i], stride_[i], true));
+    imgSizeH_.push_back(imageSize(
+        outputH_[i], filterSizeY_[i], paddingY_[i], strideY_[i], true));
+    imgSizeD_.push_back(imageSize(
+        outputD_[i], filterSizeZ_[i], paddingZ_[i], strideZ_[i], true));
+    NOut_.push_back(imgSizeD_[i] * imgSizeH_[i] * imgSizeW_[i]);
+    N_.push_back(outputD_[i] * outputH_[i] * outputW_[i]);
     CHECK(layerSize == 0 || N_[i] * size_t(numFilters_) == layerSize);
     layerSize += NOut_[i] * numFilters_;
   }
-  getOutput().setFrameHeight(outputH_[0]);
-  getOutput().setFrameWidth(outputW_[0]);
-  getOutput().setFrameDepth(outputD_[0]);
+  getOutput().setFrameHeight(imgSizeH_[0]);
+  getOutput().setFrameWidth(imgSizeW_[0]);
+  getOutput().setFrameDepth(imgSizeD_[0]);
   return layerSize;
 }
 
@@ -103,9 +103,9 @@ void DeConv3DLayer::forward(PassType passType) {
       }
       colBuf_->col2Vol(outMat->getData() + n * outMat->getStride(),
                        numFilters_,
-                       outputD_[i],
-                       outputH_[i],
-                       outputW_[i],
+                       imgSizeD_[i],
+                       imgSizeH_[i],
+                       imgSizeW_[i],
                        filterSizeZ_[i],
                        filterSizeY_[i],
                        filterSize_[i],
@@ -144,9 +144,9 @@ void DeConv3DLayer::backward(const UpdateCallback &callback) {
         colBuf_->vol2Col(
             getOutputGrad()->getData() + n * getOutputGrad()->getStride(),
             numFilters_,
-            outputD_[i],
-            outputH_[i],
-            outputW_[i],
+            imgSizeD_[i],
+            imgSizeH_[i],
+            imgSizeW_[i],
             filterSizeZ_[i],
             filterSizeY_[i],
             filterSize_[i],
diff --git a/paddle/gserver/layers/ExpandConvBaseLayer.cpp b/paddle/gserver/layers/ExpandConvBaseLayer.cpp
deleted file mode 100644
index 2b7bef0a75..0000000000
--- a/paddle/gserver/layers/ExpandConvBaseLayer.cpp
+++ /dev/null
@@ -1,124 +0,0 @@
-/* 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 "ExpandConvBaseLayer.h"
-
-#include "paddle/utils/Logging.h"
-namespace paddle {
-
-bool ExpandConvBaseLayer::init(const LayerMap &layerMap,
-                               const ParameterMap &parameterMap) {
-  /* Initialize the basic convolutional parent class */
-  ConvBaseLayer::init(layerMap, parameterMap);
-
-  int index = 0;
-  for (auto &inputConfig : config_.inputs()) {
-    const ConvConfig &conf = inputConfig.conv_conf();
-    /* Consistent caffe mode for multiple input */
-    caffeMode_ = conf.caffe_mode();
-
-    // create a new weight
-    size_t height, width;
-    height = filterPixels_[index] * filterChannels_[index];
-    width = (!isDeconv_) ? numFilters_ : channels_[index];
-    CHECK_EQ(parameters_[index]->getSize(), width * height);
-    Weight *w = new Weight(height, width, parameters_[index]);
-    weights_.emplace_back(w);
-    index++;
-  }
-  if (biasParameter_.get()) {
-    if (sharedBiases_) {
-      CHECK_EQ((size_t)numFilters_, biasParameter_->getSize());
-      biases_ =
-          std::unique_ptr<Weight>(new Weight(numFilters_, 1, biasParameter_));
-    } else {
-      biases_ =
-          std::unique_ptr<Weight>(new Weight(getSize(), 1, biasParameter_));
-    }
-  }
-  getOutputSize();
-
-  return true;
-}
-
-size_t ExpandConvBaseLayer::getOutputSize() {
-  CHECK_NE(inputLayers_.size(), 0UL);
-  size_t layerSize = ConvBaseLayer::calOutputSize();
-  return layerSize;
-}
-
-void ExpandConvBaseLayer::addSharedBias() {
-  size_t mapW = getOutputSize() / numFilters_;
-  size_t mapH = getOutputValue()->getElementCnt() / mapW;
-  MatrixPtr out =
-      Matrix::create(getOutputValue()->getData(), mapH, mapW, false, useGpu_);
-
-  Matrix::resizeOrCreate(transOutValue_, mapW, mapH, false, useGpu_);
-
-  out->transpose(transOutValue_, false);  // false means no memory allocation
-  transOutValue_->reshape(transOutValue_->getElementCnt() / numFilters_,
-                          numFilters_);
-
-  MatrixPtr bias = Matrix::create(biases_->getW()->getData(),
-                                  1,
-                                  biases_->getW()->getElementCnt(),
-                                  false,
-                                  useGpu_);
-  transOutValue_->addBias(*bias, 1.0f);
-
-  transOutValue_->reshape(mapW, mapH);
-  transOutValue_->transpose(out, false);  // false means no memory allocation
-
-  out->clear();
-  bias->clear();
-}
-
-void ExpandConvBaseLayer::addUnsharedBias() {
-  MatrixPtr outValue = getOutputValue();
-  MatrixPtr bias = Matrix::create(biases_->getW()->getData(),
-                                  1,
-                                  biases_->getW()->getElementCnt(),
-                                  false,
-                                  useGpu_);
-  outValue->addBias(*bias, 1.0f);
-}
-
-void ExpandConvBaseLayer::bpropSharedBias(MatrixPtr biases, MatrixPtr v) {
-  size_t mapW = getOutputSize() / numFilters_;
-  size_t mapH = v->getElementCnt() / mapW;
-  MatrixPtr vTmp = Matrix::create(v->getData(), mapH, mapW, false, useGpu_);
-
-  Matrix::resizeOrCreate(transOutValue_, mapW, mapH, false, useGpu_);
-
-  vTmp->transpose(transOutValue_, false);  // false means no memory allocation
-  transOutValue_->reshape(transOutValue_->getElementCnt() / numFilters_,
-                          numFilters_);
-  biases->collectBias(*transOutValue_, 1.0f);
-}
-
-void ExpandConvBaseLayer::bpropBiases(MatrixPtr v) {
-  MatrixPtr biases = Matrix::create(biases_->getWGrad()->getData(),
-                                    1,
-                                    biases_->getWGrad()->getElementCnt(),
-                                    false,
-                                    useGpu_);
-  if (sharedBiases_) {
-    bpropSharedBias(biases, v);
-  } else {
-    biases->collectBias(*v, 1.0f);
-  }
-  biases->clear();
-}
-
-}  // namespace paddle
diff --git a/paddle/gserver/layers/ExpandConvBaseLayer.h b/paddle/gserver/layers/ExpandConvBaseLayer.h
deleted file mode 100644
index 01c699d234..0000000000
--- a/paddle/gserver/layers/ExpandConvBaseLayer.h
+++ /dev/null
@@ -1,57 +0,0 @@
-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License. */
-
-#pragma once
-
-#include <vector>
-#include "ConvBaseLayer.h"
-#include "paddle/math/Matrix.h"
-
-namespace paddle {
-
-/**
- * @brief A subclass of ConvBaseLayer that is a superclass of both
- * ExpandConvLayer and ExpandConvTransLayer
- */
-class ExpandConvBaseLayer : public ConvBaseLayer {
-protected:
-  /// The transpose of output, which is an auxiliary matrix.
-  MatrixPtr transOutValue_;
-
-public:
-  explicit ExpandConvBaseLayer(const LayerConfig& config)
-      : ConvBaseLayer(config) {}
-
-  ~ExpandConvBaseLayer() {}
-
-  bool init(const LayerMap& layerMap,
-            const ParameterMap& parameterMap) override;
-
-  size_t getOutputSize();
-
-  /**
-   * Add shared bias.
-   */
-  void addSharedBias();
-
-  /**
-   * Add unshared bias.
-   */
-  void addUnsharedBias();
-
-  void bpropSharedBias(MatrixPtr biases, MatrixPtr v);
-  void bpropBiases(MatrixPtr v);
-};
-
-}  // namespace paddle
diff --git a/paddle/gserver/layers/ExpandConvLayer.cpp b/paddle/gserver/layers/ExpandConvLayer.cpp
index 20de475fc3..48dfcb49a4 100644
--- a/paddle/gserver/layers/ExpandConvLayer.cpp
+++ b/paddle/gserver/layers/ExpandConvLayer.cpp
@@ -36,7 +36,36 @@ inline bool isDepthwiseConv(int channels, int groups) {
 bool ExpandConvLayer::init(const LayerMap &layerMap,
                            const ParameterMap &parameterMap) {
   /* Initialize the basic convolutional parent class */
-  ExpandConvBaseLayer::init(layerMap, parameterMap);
+  ConvBaseLayer::init(layerMap, parameterMap);
+
+  int index = 0;
+  for (auto &inputConfig : config_.inputs()) {
+    const ConvConfig &conf = inputConfig.conv_conf();
+    /* Consistent caffe mode for multiple input */
+    caffeMode_ = conf.caffe_mode();
+
+    // create a new weight
+    size_t height, width;
+    height = filterPixels_[index] * filterChannels_[index];
+    width = (!isDeconv_) ? numFilters_ : channels_[index];
+    CHECK_EQ(parameters_[index]->getSize(), width * height);
+    Weight *w = new Weight(height, width, parameters_[index]);
+    weights_.emplace_back(w);
+    index++;
+  }
+
+  if (biasParameter_.get()) {
+    if (sharedBiases_) {
+      CHECK_EQ((size_t)numFilters_, biasParameter_->getSize());
+      biases_ = std::unique_ptr<Weight>(
+          new Weight(1, numFilters_, biasParameter_, 0));
+    } else {
+      biases_ =
+          std::unique_ptr<Weight>(new Weight(1, getSize(), biasParameter_, 0));
+    }
+  }
+
+  getOutputSize();
 
   size_t numInputs = config_.inputs_size();
   inputShape_.resize(numInputs);
@@ -108,6 +137,12 @@ bool ExpandConvLayer::init(const LayerMap &layerMap,
   return true;
 }
 
+size_t ExpandConvLayer::getOutputSize() {
+  CHECK_NE(inputLayers_.size(), 0UL);
+  size_t layerSize = ConvBaseLayer::calOutputSize();
+  return layerSize;
+}
+
 // i is the index of input layers
 #define BACKWARD_INPUT(i, inputs, outputs) \
   backward_[2 * i]->calc(inputs, outputs)
@@ -155,11 +190,7 @@ void ExpandConvLayer::forward(PassType passType) {
 
   /* add the bias-vector */
   if (biases_.get()) {
-    if (sharedBiases_) {
-      addSharedBias();
-    } else {
-      addUnsharedBias();
-    }
+    output_.value->addBias(*biases_->getW(), 1.0, sharedBiases_);
   }
 
   /* activation */
@@ -171,7 +202,7 @@ void ExpandConvLayer::backward(const UpdateCallback &callback) {
 
   MatrixPtr outGrad = getOutputGrad();
   if (biases_ && biases_->getWGrad()) {
-    bpropBiases(outGrad);
+    biases_->getWGrad()->collectBias(*getOutputGrad(), 1, sharedBiases_);
     /* Increasing the number of gradient */
     biases_->getParameterPtr()->incUpdate(callback);
   }
diff --git a/paddle/gserver/layers/ExpandConvLayer.h b/paddle/gserver/layers/ExpandConvLayer.h
index a1f943d152..a0873de192 100644
--- a/paddle/gserver/layers/ExpandConvLayer.h
+++ b/paddle/gserver/layers/ExpandConvLayer.h
@@ -15,7 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include <vector>
-#include "ExpandConvBaseLayer.h"
+#include "ConvBaseLayer.h"
 #include "paddle/math/Matrix.h"
 
 namespace paddle {
@@ -28,10 +28,9 @@ namespace paddle {
  * The config file api is img_conv_layer.
  */
 
-class ExpandConvLayer : public ExpandConvBaseLayer {
+class ExpandConvLayer : public ConvBaseLayer {
 public:
-  explicit ExpandConvLayer(const LayerConfig& config)
-      : ExpandConvBaseLayer(config) {}
+  explicit ExpandConvLayer(const LayerConfig& config) : ConvBaseLayer(config) {}
 
   ~ExpandConvLayer() {}
 
@@ -41,6 +40,8 @@ public:
   void forward(PassType passType) override;
   void backward(const UpdateCallback& callback) override;
 
+  size_t getOutputSize();
+
 protected:
   std::vector<TensorShape> inputShape_;
   std::vector<TensorShape> filterShape_;
diff --git a/paddle/gserver/layers/Layer.h b/paddle/gserver/layers/Layer.h
index edef36194a..4002a3d074 100644
--- a/paddle/gserver/layers/Layer.h
+++ b/paddle/gserver/layers/Layer.h
@@ -49,6 +49,12 @@ struct LayerState {
 };
 typedef std::shared_ptr<LayerState> LayerStatePtr;
 
+/// Paddle device ID, MKLDNN is -2, CPU is -1
+enum PADDLE_DEVICE_ID {
+  MKLDNN_DEVICE = -2,
+  CPU_DEVICE = -1,
+};
+
 /**
  * @brief Base class for layer.
  * Define necessary variables and functions for every layer.
@@ -59,11 +65,6 @@ protected:
   LayerConfig config_;
   /// whether to use GPU
   bool useGpu_;
-  /// Paddle device ID, MKLDNN is -2, CPU is -1
-  enum PADDLE_DEVICE_ID {
-    MKLDNN_DEVICE = -2,
-    CPU_DEVICE = -1,
-  };
   /// Device Id. MKLDNN is -2, CPU is -1, and GPU is 0, 1, 2 ...
   int deviceId_;
   /// Input layers
diff --git a/paddle/gserver/layers/MKLDNNConvLayer.cpp b/paddle/gserver/layers/MKLDNNConvLayer.cpp
new file mode 100644
index 0000000000..9088744bee
--- /dev/null
+++ b/paddle/gserver/layers/MKLDNNConvLayer.cpp
@@ -0,0 +1,544 @@
+/* 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 "MKLDNNConvLayer.h"
+#include "paddle/math/MathUtils.h"
+#include "paddle/utils/Logging.h"
+
+using namespace mkldnn;  // NOLINT
+typedef memory::format format;
+
+namespace paddle {
+
+REGISTER_LAYER(mkldnn_conv, MKLDNNConvLayer);
+
+bool MKLDNNConvLayer::init(const LayerMap& layerMap,
+                           const ParameterMap& parameterMap) {
+  if (!MKLDNNLayer::init(layerMap, parameterMap)) {
+    return false;
+  }
+  CHECK_EQ(inputLayers_.size(), 1) << "Only support one input layer yet";
+  CHECK_EQ(inputLayers_.size(), parameters_.size());
+  CHECK(config_.shared_biases()) << "Only support shared biases yet";
+
+  oc_ = config_.num_filters();
+  const ConvConfig& conf = config_.inputs(0).conv_conf();
+  ic_ = conf.channels();
+  fw_ = conf.filter_size();
+  fh_ = conf.filter_size_y();
+  pw_ = conf.padding();
+  ph_ = conf.padding_y();
+  dw_ = conf.dilation();
+  dh_ = conf.dilation_y();
+  sw_ = conf.stride();
+  sh_ = conf.stride_y();
+  gp_ = conf.groups();
+  oh_ = conf.output_y();
+  ow_ = conf.output_x();
+  ih_ = conf.img_size_y();
+  iw_ = conf.img_size();
+  caffeMode_ = conf.caffe_mode();
+  CHECK(caffeMode_) << "Only support caffe mode yet";
+  CHECK(dh_ == 1 && dw_ == 1) << "Only support dilation 1 yet";
+  // check group setting
+  CHECK_EQ((oc_ / gp_) * gp_, oc_) << "group is indivisible for oc";
+  CHECK_EQ((ic_ / gp_) * gp_, ic_) << "group is indivisible for ic";
+
+  // create weight
+  size_t height = oc_ / gp_;
+  size_t width = ic_ * fh_ * fw_;
+  CHECK_EQ(parameters_[0]->getSize(), height * width);
+  weight_ =
+      std::unique_ptr<Weight>(new Weight(height, width, parameters_[0], 0));
+
+  // create biases
+  if (biasParameter_.get() != NULL) {
+    biases_ = std::unique_ptr<Weight>(new Weight(1, oc_, biasParameter_));
+  }
+  return true;
+}
+
+void MKLDNNConvLayer::convertWeightsFromPaddle() {
+  if (hasInitedWgt_) {
+    return;
+  }
+
+  CHECK(wgtVal_) << "should have been initialized";
+  // the paddle weight format is oihw or goihw
+  auto targetDim = wgtVal_->getDims();
+  auto srcFmt = (gp_ == 1) ? memory::format::oihw : memory::format::goihw;
+  wgtVal_->reorderDataFrom(wgtVal_, srcFmt, targetDim);
+  hasInitedWgt_ = true;
+}
+
+void MKLDNNConvLayer::convertWeightsToPaddle() {
+  CHECK(wgtVal_) << "should have been initialized";
+  auto targetDim = wgtVal_->getDims();
+  auto dstFmt = (gp_ == 1) ? memory::format::oihw : memory::format::goihw;
+  wgtVal_->reorderDataTo(wgtVal_, dstFmt, targetDim);
+}
+
+void MKLDNNConvLayer::reshape(
+    int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) {
+  reshapeInput(bs, ih, iw);
+
+  // cal output sizes
+  // oc can not be changed
+  int fh = (fh_ - 1) * dh_ + 1;
+  int fw = (fw_ - 1) * dw_ + 1;
+  oh = outputSize(ih, fh, ph_, sh_, caffeMode_);
+  ow = outputSize(iw, fw, pw_, sw_, caffeMode_);
+
+  reshapeOutput(oh, ow);
+  resizeOutput(bs, oc * oh * ow);
+
+  printSizeInfo();
+}
+
+void MKLDNNConvLayer::resetFwd(std::vector<primitive>& pipeline,
+                               MKLDNNMatrixPtr& in,
+                               MKLDNNMatrixPtr& wgt,
+                               MKLDNNMatrixPtr& bias,
+                               MKLDNNMatrixPtr& out) {
+  resetFwdPD(fwdPD_);
+
+  resetFwdBuffers(fwdPD_, in, wgt, bias, out);
+
+  resetFwdPipeline(pipeline, fwdPD_, in, wgt, bias, out);
+
+  printValueFormatFlow();
+}
+
+void MKLDNNConvLayer::resetBwd(std::vector<primitive>& pipeline,
+                               MKLDNNMatrixPtr& in,
+                               MKLDNNMatrixPtr& wgt,
+                               MKLDNNMatrixPtr& bias,
+                               MKLDNNMatrixPtr& out) {
+  std::shared_ptr<conv_bwdWgt::primitive_desc> bwdWgtPD;
+  std::shared_ptr<conv_bwdData::primitive_desc> bwdDataPD;
+
+  resetBwdWgtPD(bwdWgtPD);
+
+  resetBwdDataPD(bwdDataPD);
+
+  resetBwdBuffers(bwdWgtPD, bwdDataPD, in, wgt, bias, out);
+
+  resetBwdPipeline(pipeline, bwdWgtPD, bwdDataPD, in, wgt, bias, out);
+
+  printGradFormatFlow();
+}
+
+void MKLDNNConvLayer::updateInputData() {
+  cpuInVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
+}
+
+void MKLDNNConvLayer::updateWeights(const UpdateCallback& callback) {
+  weight_->getParameterPtr()->incUpdate(callback);
+  if (biases_ && biases_->getWGrad()) {
+    biases_->getParameterPtr()->incUpdate(callback);
+  }
+}
+
+void MKLDNNConvLayer::loadConvSettings(memory::dims& wgt,
+                                       memory::dims& bias,
+                                       memory::dims& stride,
+                                       memory::dims& dilation,
+                                       memory::dims& padL,
+                                       memory::dims& padR) {
+  wgt = (gp_ == 1) ? memory::dims{oc_, ic_, fh_, fw_}
+                   : memory::dims{gp_, oc_ / gp_, ic_ / gp_, fh_, fw_};
+  bias = memory::dims{oc_};
+  stride = memory::dims{sh_, sw_};
+  padL = memory::dims{ph_, pw_};
+  padR = getPaddingR();
+  // note: mkldnn dilation start from 0
+  dilation = memory::dims{dh_ - 1, dw_ - 1};
+}
+
+void MKLDNNConvLayer::resetFwdPD(
+    std::shared_ptr<conv_fwd::primitive_desc>& pd) {
+  // dims for conv
+  memory::dims inDims = memory::dims{bs_, ic_, ih_, iw_};
+  memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
+  memory::dims wgtDims, biasDims, strides, dilations, padL, padR;
+  loadConvSettings(wgtDims, biasDims, strides, dilations, padL, padR);
+
+  prop_kind pk = passType_ == PASS_TEST ? prop_kind::forward_scoring
+                                        : prop_kind::forward_training;
+  algorithm algo = algorithm::convolution_direct;
+  padding_kind padKind = padding_kind::zero;
+  conv_fwd::desc fwdDesc =
+      biases_ && biases_->getW()
+          ? conv_fwd::desc(pk,
+                           algo,
+                           MKLDNNMatrix::createMemoryDesc(inDims),
+                           MKLDNNMatrix::createMemoryDesc(wgtDims),
+                           MKLDNNMatrix::createMemoryDesc(biasDims),
+                           MKLDNNMatrix::createMemoryDesc(outDims),
+                           strides,
+                           dilations,
+                           padL,
+                           padR,
+                           padKind)
+          : conv_fwd::desc(pk,
+                           algo,
+                           MKLDNNMatrix::createMemoryDesc(inDims),
+                           MKLDNNMatrix::createMemoryDesc(wgtDims),
+                           MKLDNNMatrix::createMemoryDesc(outDims),
+                           strides,
+                           dilations,
+                           padL,
+                           padR,
+                           padKind);
+  pd.reset(new conv_fwd::primitive_desc(fwdDesc, engine_));
+}
+
+void MKLDNNConvLayer::resetFwdBuffers(
+    std::shared_ptr<conv_fwd::primitive_desc>& pd,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias,
+    MKLDNNMatrixPtr& out) {
+  CHECK(pd);
+  resetInValue(pd, in);
+
+  resetWgtBiasValue(pd, wgt, bias);
+
+  resetOutValue(pd, out);
+}
+
+void MKLDNNConvLayer::resetFwdPipeline(
+    std::vector<primitive>& pipeline,
+    std::shared_ptr<conv_fwd::primitive_desc>& pd,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias,
+    MKLDNNMatrixPtr& out) {
+  pipeline.clear();
+
+  if (cvtInVal_) {
+    pipeline.push_back(*cvtInVal_);
+  }
+
+  if (bias) {
+    fwd_.reset(new conv_fwd(*pd, *in, *wgt, *bias, *out));
+  } else {
+    fwd_.reset(new conv_fwd(*pd, *in, *wgt, *out));
+  }
+  pipeline.push_back(*fwd_);
+
+  if (cvtOutVal_) {
+    pipeline.push_back(*cvtOutVal_);
+  }
+}
+
+void MKLDNNConvLayer::resetInValue(
+    std::shared_ptr<conv_fwd::primitive_desc>& pd, MKLDNNMatrixPtr& in) {
+  const MatrixPtr& inMat = inputLayers_[0]->getOutput().value;
+  in = MKLDNNMatrix::create(inMat, pd->src_primitive_desc());
+
+  // create buffer and reorder if input value do not match
+  cpuInVal_ = nullptr;
+  cvtInVal_ = nullptr;
+  if (inputIsOnlyMKLDNN()) {
+    MKLDNNMatrixPtr dnnIn = std::dynamic_pointer_cast<MKLDNNMatrix>(inMat);
+    CHECK(dnnIn) << "Input should be MKLDNNMatrix";
+    if (dnnIn->getPrimitiveDesc() != in->getPrimitiveDesc()) {
+      CHECK_EQ(dnnIn->getFormat(), format::nc);
+      CHECK(ih_ == 1 && iw_ == 1) << "when input is nc format";
+      // create a new one with nchw format and same data
+      memory::dims inDims = memory::dims{bs_, ic_, 1, 1};
+      dnnIn = MKLDNNMatrix::create(inMat, inDims, format::nchw, engine_);
+      CHECK(dnnIn->getPrimitiveDesc() == in->getPrimitiveDesc());
+    }
+    in = dnnIn;
+  } else {
+    const MatrixPtr& cpuIn = getInputValue(0, CPU_DEVICE);
+    memory::dims inDims = memory::dims{bs_, ic_, ih_, iw_};
+    cpuInVal_ = MKLDNNMatrix::create(cpuIn, inDims, format::nchw, engine_);
+    if (cpuInVal_->getPrimitiveDesc() != in->getPrimitiveDesc()) {
+      // create new mkldnn matrix
+      in = MKLDNNMatrix::create(nullptr, pd->src_primitive_desc());
+      cvtInVal_ = MKLDNNMatrix::createReorder(cpuInVal_, in);
+      CHECK(cvtInVal_) << "should not be emptry";
+    } else {
+      in = cpuInVal_;
+    }
+  }
+}
+
+void MKLDNNConvLayer::resetWgtBiasValue(
+    std::shared_ptr<conv_fwd::primitive_desc>& pd,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias) {
+  wgt = MKLDNNMatrix::create(weight_->getW(), pd->weights_primitive_desc());
+  VLOG(MKLDNN_FMTS) << "Weight value format: " << wgt->getFormat();
+
+  bias = (biases_ && biases_->getW())
+             ? MKLDNNMatrix::create(biases_->getW(), pd->bias_primitive_desc())
+             : nullptr;
+}
+
+void MKLDNNConvLayer::resetOutValue(
+    std::shared_ptr<conv_fwd::primitive_desc>& pd, MKLDNNMatrixPtr& out) {
+  out = MKLDNNMatrix::create(output_.value, pd->dst_primitive_desc());
+
+  // change original output value from cpu matrix to mkldnn matrix
+  output_.value = std::dynamic_pointer_cast<Matrix>(out);
+
+  // create reorder if output value has cpu device and pd do not match
+  cpuOutVal_ = nullptr;
+  cpuOutVal_ = nullptr;
+  if (!outputIsOnlyMKLDNN()) {
+    const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).value;
+    memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
+    cpuOutVal_ = MKLDNNMatrix::create(cpuOut, outDims, format::nchw, engine_);
+    if (cpuOutVal_->getPrimitiveDesc() != out->getPrimitiveDesc()) {
+      cvtOutVal_ = MKLDNNMatrix::createReorder(out, cpuOutVal_);
+      CHECK(cvtOutVal_) << "should not be emptry";
+    } else {
+      // CPU output share the same data of MKLDNN output
+      cpuOut->setData(out->getData());
+      cpuOutVal_ = out;
+    }
+  }
+}
+
+void MKLDNNConvLayer::resetBwdWgtPD(
+    std::shared_ptr<conv_bwdWgt::primitive_desc>& pd) {
+  memory::dims wgtDims, biasDims, strides, dilations, padL, padR;
+  loadConvSettings(wgtDims, biasDims, strides, dilations, padL, padR);
+
+  // create backward weight using input, output and weight value memory desc
+  CHECK(inVal_) << "Should have input value";
+  CHECK(outVal_) << "Should have output value";
+  CHECK(wgtVal_) << "Should have weight value";
+  algorithm algo = algorithm::convolution_direct;
+  padding_kind padKind = padding_kind::zero;
+  auto bwdWgtDesc = biasVal_ != nullptr
+                        ? conv_bwdWgt::desc(algo,
+                                            inVal_->getMemoryDesc(),
+                                            wgtVal_->getMemoryDesc(),
+                                            biasVal_->getMemoryDesc(),
+                                            outVal_->getMemoryDesc(),
+                                            strides,
+                                            padL,
+                                            padR,
+                                            padKind)
+                        : conv_bwdWgt::desc(algo,
+                                            inVal_->getMemoryDesc(),
+                                            wgtVal_->getMemoryDesc(),
+                                            outVal_->getMemoryDesc(),
+                                            strides,
+                                            padL,
+                                            padR,
+                                            padKind);
+  pd.reset(new conv_bwdWgt::primitive_desc(bwdWgtDesc, engine_, *fwdPD_));
+  CHECK(pd->src_primitive_desc() == inVal_->getPrimitiveDesc())
+      << "primitive desc of in value should equal";
+  CHECK(pd->diff_dst_primitive_desc() == outVal_->getPrimitiveDesc())
+      << "primitive desc of out grad should equal the out value";
+  CHECK(pd->diff_weights_primitive_desc() == wgtVal_->getPrimitiveDesc())
+      << "primitive desc of weight grad should equal the weight value";
+}
+
+void MKLDNNConvLayer::resetBwdDataPD(
+    std::shared_ptr<conv_bwdData::primitive_desc>& pd) {
+  pd = nullptr;
+  if (inputLayers_[0]->getOutput().grad == nullptr) {
+    return;
+  }
+
+  memory::dims wgtDims, biasDims, strides, dilations, padL, padR;
+  loadConvSettings(wgtDims, biasDims, strides, dilations, padL, padR);
+  CHECK(inVal_) << "Should have input value";
+  CHECK(outVal_) << "Should have output value";
+  // create backward data using input and output value memory desc
+  // but using weight memory desc with any format
+  auto bwdDataDesc = conv_bwdData::desc(algorithm::convolution_direct,
+                                        inVal_->getMemoryDesc(),
+                                        MKLDNNMatrix::createMemoryDesc(wgtDims),
+                                        outVal_->getMemoryDesc(),
+                                        strides,
+                                        padL,
+                                        padR,
+                                        padding_kind::zero);
+  pd.reset(new conv_bwdData::primitive_desc(bwdDataDesc, engine_, *fwdPD_));
+  CHECK(pd->diff_src_primitive_desc() == inVal_->getPrimitiveDesc())
+      << "primitive desc of in grad should equal the in value";
+  CHECK(pd->diff_dst_primitive_desc() == outVal_->getPrimitiveDesc())
+      << "primitive desc of out grad should equal";
+}
+
+void MKLDNNConvLayer::resetBwdBuffers(
+    std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
+    std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias,
+    MKLDNNMatrixPtr& out) {
+  CHECK(wgtPD);
+  resetOutGrad(wgtPD, out);
+
+  resetWgtBiasGrad(wgtPD, wgt, bias);
+
+  resetInGrad(dataPD, in);
+
+  resetWgtValBwdData(dataPD, wgtValBwdData_);
+}
+
+void MKLDNNConvLayer::resetBwdPipeline(
+    std::vector<primitive>& pipeline,
+    std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
+    std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias,
+    MKLDNNMatrixPtr& out) {
+  pipeline.clear();
+
+  if (cvtOutGrad_) {
+    pipeline.push_back(*cvtOutGrad_);
+  }
+
+  // add bwdWgt handle
+  if (bias) {
+    bwdWgt_.reset(new conv_bwdWgt(*wgtPD, *inVal_, *out, *wgt, *bias));
+  } else {
+    bwdWgt_.reset(new conv_bwdWgt(*wgtPD, *inVal_, *out, *wgt));
+  }
+  pipeline.push_back(*bwdWgt_);
+
+  if (dataPD == nullptr) {
+    return;
+  }
+
+  if (cvtWgtVal_) {
+    pipeline.push_back(*cvtWgtVal_);
+  }
+
+  // add bwdData handle
+  CHECK(wgtValBwdData_) << "Should have weight memory";
+  bwdData_.reset(new conv_bwdData(*dataPD, *out, *wgtValBwdData_, *in));
+  pipeline.push_back(*bwdData_);
+
+  if (cvtInGrad_) {
+    pipeline.push_back(*cvtInGrad_);
+  }
+}
+
+void MKLDNNConvLayer::resetOutGrad(
+    std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD, MKLDNNMatrixPtr& out) {
+  const MatrixPtr& outMat = output_.grad;
+  out = MKLDNNMatrix::create(outMat, wgtPD->diff_dst_primitive_desc());
+  CHECK(outVal_ != nullptr &&
+        out->getPrimitiveDesc() == outVal_->getPrimitiveDesc())
+      << "primitive desc of out grad and value should be equal";
+
+  // TODO(TJ): merge outgrad
+  // create reorder if has output grad does not match
+  cpuOutGrad_ = nullptr;
+  cvtOutGrad_ = nullptr;
+  if (!outputIsOnlyMKLDNN()) {
+    const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).grad;
+    // same PrimitiveDesc with cpuInVal_
+    CHECK(cpuOutVal_);
+    cpuOutGrad_ = MKLDNNMatrix::create(cpuOut, cpuOutVal_->getPrimitiveDesc());
+    if (cpuOutGrad_->getPrimitiveDesc() == out->getPrimitiveDesc()) {
+      outMat->setData(cpuOut->getData());
+      out = cpuOutGrad_;
+    } else {
+      cvtOutGrad_ = MKLDNNMatrix::createReorder(cpuOutGrad_, out);
+      CHECK(cvtOutGrad_);
+    }
+  }
+}
+
+void MKLDNNConvLayer::resetWgtBiasGrad(
+    std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias) {
+  wgt = MKLDNNMatrix::create(weight_->getWGrad(),
+                             wgtPD->diff_weights_primitive_desc());
+  CHECK(nullptr != wgtVal_ &&
+        wgt->getPrimitiveDesc() == wgtVal_->getPrimitiveDesc())
+      << "primitive desc of weight grad and value should be equal";
+  VLOG(MKLDNN_FMTS) << "weight grad format: " << wgt->getFormat();
+
+  bias = nullptr;
+  if (biasVal_ == nullptr) {
+    return;
+  }
+  bias = MKLDNNMatrix::create(biases_->getWGrad(),
+                              wgtPD->diff_bias_primitive_desc());
+  CHECK(bias->getPrimitiveDesc() == biasVal_->getPrimitiveDesc())
+      << "primitive desc of bias grad should equal the bias value";
+}
+
+void MKLDNNConvLayer::resetInGrad(
+    std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
+    MKLDNNMatrixPtr& in) {
+  if (dataPD == nullptr) {
+    return;
+  }
+
+  // TODO(TJ): use outputMaps_ ways to get the inGrad_ when merge outgrad done
+  in = MKLDNNMatrix::create(inputLayers_[0]->getOutput().grad,
+                            dataPD->diff_src_primitive_desc());
+  CHECK(nullptr != inVal_ &&
+        in->getPrimitiveDesc() == inVal_->getPrimitiveDesc())
+      << "primitive desc of input grad and value should be equal";
+
+  // create reorder if has output grad does not match
+  cpuInGrad_ = nullptr;
+  cvtInGrad_ = nullptr;
+  if (!inputIsOnlyMKLDNN()) {
+    const MatrixPtr& cpuIn = getInputGrad(0, CPU_DEVICE);
+    // same PrimitiveDesc with cpuInVal_
+    CHECK(cpuInVal_);
+    cpuInGrad_ = MKLDNNMatrix::create(cpuIn, cpuInVal_->getPrimitiveDesc());
+    if (cpuInGrad_->getPrimitiveDesc() != in->getPrimitiveDesc()) {
+      const MatrixPtr& dnnIn = getInputGrad(0, MKLDNN_DEVICE);
+      in = MKLDNNMatrix::create(dnnIn, in->getPrimitiveDesc());
+      cvtInGrad_ = MKLDNNMatrix::createReorder(in, cpuInGrad_);
+      CHECK(cvtInGrad_);
+    } else {
+      in = cpuInGrad_;
+    }
+  }
+}
+
+void MKLDNNConvLayer::resetWgtValBwdData(
+    std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
+    MKLDNNMatrixPtr& wgt) {
+  if (dataPD == nullptr) {
+    return;
+  }
+
+  // create new weight value for backward data, and create reorder if necessary
+  // since the primitive_desc would be different with wgtVal_
+  CHECK(wgtVal_) << "should have weight value";
+  if (dataPD->weights_primitive_desc() != wgtVal_->getPrimitiveDesc()) {
+    wgtValBwdData_ =
+        MKLDNNMatrix::create(nullptr, dataPD->weights_primitive_desc());
+    cvtWgtVal_ = MKLDNNMatrix::createReorder(wgtVal_, wgtValBwdData_);
+    CHECK(cvtWgtVal_);
+  } else {
+    wgtValBwdData_ = wgtVal_;
+  }
+  VLOG(MKLDNN_FMTS) << "weight value format for backward data"
+                    << wgtValBwdData_->getFormat();
+}
+
+}  // namespace paddle
diff --git a/paddle/gserver/layers/MKLDNNConvLayer.h b/paddle/gserver/layers/MKLDNNConvLayer.h
new file mode 100644
index 0000000000..f84f2f737c
--- /dev/null
+++ b/paddle/gserver/layers/MKLDNNConvLayer.h
@@ -0,0 +1,253 @@
+/* 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 {
+typedef mkldnn::convolution_forward conv_fwd;
+typedef mkldnn::convolution_backward_weights conv_bwdWgt;
+typedef mkldnn::convolution_backward_data conv_bwdData;
+
+/**
+ * @brief A subclass of MKLDNNLayer conv layer.
+ *
+ * The config file api is mkldnn_conv
+ */
+class MKLDNNConvLayer : public MKLDNNLayer {
+protected:
+  // padding height and width
+  int ph_, pw_;
+  // stride height and width
+  int sh_, sw_;
+  // dilation height and width
+  int dh_, dw_;
+  // filter(kenerl) height and width
+  int fh_, fw_;
+  // group number
+  int gp_;
+
+  // in resetBwdData, the format of wgtValBwdData_ is different with wgtVal_
+  MKLDNNMatrixPtr wgtValBwdData_;
+  // convert handle from wgtVal_ to wgtValBwdData_
+  std::shared_ptr<mkldnn::reorder> cvtWgtVal_;
+
+  // save forward primitive_desc, which can be used backward
+  std::shared_ptr<conv_fwd::primitive_desc> fwdPD_;
+
+  // MKLDNNMatrixPtr which should be created from CPU Device
+  MKLDNNMatrixPtr cpuInVal_;
+  MKLDNNMatrixPtr cpuInGrad_;
+  MKLDNNMatrixPtr cpuOutVal_;
+  MKLDNNMatrixPtr cpuOutGrad_;
+  // convert handle between CPU device and MKLDNN device
+  std::shared_ptr<mkldnn::reorder> cvtInVal_;
+  std::shared_ptr<mkldnn::reorder> cvtInGrad_;
+  std::shared_ptr<mkldnn::reorder> cvtOutVal_;
+  std::shared_ptr<mkldnn::reorder> cvtOutGrad_;
+
+  // whether the weight has been init
+  bool hasInitedWgt_;
+
+  // true by default, which impact the calculation of output image size.
+  // details can refer to mathUtil.h
+  bool caffeMode_;
+
+  // weight and bias
+  std::unique_ptr<Weight> weight_;
+  std::unique_ptr<Weight> biases_;
+
+public:
+  explicit MKLDNNConvLayer(const LayerConfig& config)
+      : MKLDNNLayer(config), hasInitedWgt_(false), caffeMode_(true) {}
+
+  ~MKLDNNConvLayer() {}
+
+  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 updateInputData() override;
+
+  void updateWeights(const UpdateCallback& callback) override;
+
+  void convertWeightsFromPaddle() override;
+
+  void convertWeightsToPaddle() override;
+
+  void printSizeInfo() override {
+    MKLDNNLayer::printSizeInfo();
+    VLOG(MKLDNN_SIZES) << getName() << ": fh: " << fh_ << ", fw: " << fw_
+                       << ": ph: " << ph_ << ", pw: " << pw_ << ", sh: " << sh_
+                       << ", sw: " << sw_ << ", dh: " << dh_ << ", dw: " << dw_;
+  }
+
+  void printValueFormatFlow() override {
+    if (cpuInVal_) {
+      VLOG(MKLDNN_FMTS) << cpuInVal_->getFormat() << " >>>";
+    }
+    MKLDNNLayer::printValueFormatFlow();
+    if (cpuOutVal_) {
+      VLOG(MKLDNN_FMTS) << " >>> " << cpuOutVal_->getFormat();
+    }
+  }
+
+  void printGradFormatFlow() override {
+    if (cpuInGrad_) {
+      VLOG(MKLDNN_FMTS) << cpuInGrad_->getFormat() << " <<<";
+    }
+    MKLDNNLayer::printGradFormatFlow();
+    if (cpuOutGrad_) {
+      VLOG(MKLDNN_FMTS) << " <<< " << cpuOutGrad_->getFormat();
+    }
+  }
+
+protected:
+  /**
+   * load the dims settings of this conv
+   */
+  void loadConvSettings(mkldnn::memory::dims& wgt,
+                        mkldnn::memory::dims& bias,
+                        mkldnn::memory::dims& stride,
+                        mkldnn::memory::dims& dilation,
+                        mkldnn::memory::dims& padL,
+                        mkldnn::memory::dims& padR);
+
+  /**
+   * reset the forward primitive descriptor.
+   */
+  void resetFwdPD(std::shared_ptr<conv_fwd::primitive_desc>& pd);
+  /**
+   * reset the MKLDNNMatrix buffers used in forward.
+   */
+  void resetFwdBuffers(std::shared_ptr<conv_fwd::primitive_desc>& pd,
+                       MKLDNNMatrixPtr& in,
+                       MKLDNNMatrixPtr& wgt,
+                       MKLDNNMatrixPtr& bias,
+                       MKLDNNMatrixPtr& out);
+  /**
+   * reset the forward pipeline.
+   */
+  void resetFwdPipeline(std::vector<mkldnn::primitive>& pipeline,
+                        std::shared_ptr<conv_fwd::primitive_desc>& pd,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& wgt,
+                        MKLDNNMatrixPtr& bias,
+                        MKLDNNMatrixPtr& out);
+
+  /**
+   * reset MKLDNNMatrix of input value
+   */
+  void resetInValue(std::shared_ptr<conv_fwd::primitive_desc>& pd,
+                    MKLDNNMatrixPtr& in);
+  /**
+   * reset MKLDNNMatrix of weight and bias value
+   */
+  void resetWgtBiasValue(std::shared_ptr<conv_fwd::primitive_desc>& pd,
+                         MKLDNNMatrixPtr& wgt,
+                         MKLDNNMatrixPtr& bias);
+  /**
+   * reset MKLDNNMatrix of output value
+   */
+  void resetOutValue(std::shared_ptr<conv_fwd::primitive_desc>& pd,
+                     MKLDNNMatrixPtr& out);
+
+  /**
+   * reset the backward weight primitive descriptor.
+   */
+  void resetBwdWgtPD(std::shared_ptr<conv_bwdWgt::primitive_desc>& pd);
+  /**
+   * reset the backward data primitive descriptor.
+   */
+  void resetBwdDataPD(std::shared_ptr<conv_bwdData::primitive_desc>& pd);
+  /**
+   * reset the MKLDNNMatrix buffers used in backward.
+   */
+  void resetBwdBuffers(std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
+                       std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
+                       MKLDNNMatrixPtr& in,
+                       MKLDNNMatrixPtr& wgt,
+                       MKLDNNMatrixPtr& bias,
+                       MKLDNNMatrixPtr& out);
+  /**
+   * reset the backward pipeline.
+   */
+  void resetBwdPipeline(std::vector<mkldnn::primitive>& pipeline,
+                        std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
+                        std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& wgt,
+                        MKLDNNMatrixPtr& bias,
+                        MKLDNNMatrixPtr& out);
+
+  /**
+   * reset MKLDNNMatrix of output grad
+   */
+  void resetOutGrad(std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
+                    MKLDNNMatrixPtr& out);
+  /**
+   * reset MKLDNNMatrix of weight and bias grad
+   */
+  void resetWgtBiasGrad(std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
+                        MKLDNNMatrixPtr& wgt,
+                        MKLDNNMatrixPtr& bias);
+  /**
+   * reset MKLDNNMatrix of input grad
+   */
+  void resetInGrad(std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
+                   MKLDNNMatrixPtr& in);
+  /**
+   * reset MKLDNNMatrix of weight value for backward data
+   * since the primitive_desc would be different with wgtVal_
+   */
+  void resetWgtValBwdData(std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
+                          MKLDNNMatrixPtr& wgt);
+
+  /**
+   * get padding_r according to
+   * https://github.com/01org/mkl-dnn/blob/master/tests/gtests/
+   * test_convolution_forward_common.hpp
+   * @note: mkldnn dilation start from 0 while paddle start from 1
+   */
+  mkldnn::memory::dims getPaddingR() const {
+    mkldnn::memory::dims padR = {ph_, pw_};
+    for (int i = 0; i < 2; ++i) {
+      if ((ih_ - ((fh_ - 1) * dh_ + 1) + ph_ + padR[0]) / sh_ + 1 != oh_) {
+        ++padR[0];
+      }
+      if ((iw_ - ((fw_ - 1) * dw_ + 1) + pw_ + padR[1]) / sw_ + 1 != ow_) {
+        ++padR[1];
+      }
+    }
+    return padR;
+  }
+};
+
+}  // namespace paddle
diff --git a/paddle/gserver/layers/MKLDNNFcLayer.cpp b/paddle/gserver/layers/MKLDNNFcLayer.cpp
index 8318c8c519..f60e221a6e 100644
--- a/paddle/gserver/layers/MKLDNNFcLayer.cpp
+++ b/paddle/gserver/layers/MKLDNNFcLayer.cpp
@@ -14,13 +14,9 @@ limitations under the License. */
 
 #include "MKLDNNFcLayer.h"
 #include "paddle/utils/Logging.h"
-#include "paddle/utils/Stat.h"
 
 using namespace mkldnn;  // NOLINT
 typedef memory::format format;
-typedef inner_product_forward fc_fwd;
-typedef inner_product_backward_weights fc_bwdWgt;
-typedef inner_product_backward_data fc_bwdData;
 
 namespace paddle {
 
@@ -40,6 +36,8 @@ bool MKLDNNFcLayer::init(const LayerMap& layerMap,
   oc_ = getSize();
   oh_ = 1;
   ow_ = 1;
+  ih_ = 1;
+  iw_ = 1;
 
   // input size can not change in FC
   iLayerSize_ = inputLayers_[0]->getSize();
@@ -77,122 +75,163 @@ void MKLDNNFcLayer::convertWeightsToPaddle() {
   wgtVal_->reorderDataTo(wgtVal_, dstFmt, targetDim);
 }
 
-void MKLDNNFcLayer::convertOutputToOtherDevice() {
-  copyOutputInfoToOtherDevice();
-  // find other cpu device and reorder output to cpu device
-  int cnt = 0;
-  for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
-    if (outputOtherDevice_[i].deviceId == CPU_DEVICE) {
-      // fc cpu output value do not need convert
-      // just share point
-      outputOtherDevice_[i].value = output_.value;
-      ++cnt;
-    }
-  }
+void MKLDNNFcLayer::reshape(
+    int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) {
+  reshapeInput(bs, ih, iw);
 
-  if (cnt > 1) {
-    LOG(WARNING) << "should not have more than one CPU devie";
-  }
-}
-
-void MKLDNNFcLayer::reshape() {
-  const Argument& input = getInput(0, getPrev(0)->getDeviceId());
-  int batchSize = input.getBatchSize();
-  if (bs_ == batchSize) {
-    return;
-  }
-  bs_ = batchSize;
-  ih_ = input.getFrameHeight();
-  iw_ = input.getFrameWidth();
-  if (ih_ == 0) {
-    ih_ = 1;
-  }
-  if (iw_ == 0) {
-    iw_ = 1;
-  }
   CHECK_EQ(iLayerSize_, inputLayers_[0]->getSize());
-  ic_ = iLayerSize_ / (ih_ * iw_);
-  CHECK_EQ(size_t(ic_ * ih_ * iw_), iLayerSize_) << "not divisible";
-  CHECK_EQ(size_t(oc_), getSize());
+  ic = iLayerSize_ / (ih * iw);
+  CHECK_EQ(size_t(ic * ih * iw), iLayerSize_) << "not divisible";
+  CHECK_EQ(size_t(oc), getSize());
+
+  reshapeOutput(oh, ow);
+  resizeOutput(bs, oc);
+
   printSizeInfo();
+}
 
-  // reset output
-  output_.setFrameHeight(oh_);
-  output_.setFrameWidth(ow_);
-  resetOutput(bs_, oc_);
+void MKLDNNFcLayer::resetFwd(std::vector<primitive>& pipeline,
+                             MKLDNNMatrixPtr& in,
+                             MKLDNNMatrixPtr& wgt,
+                             MKLDNNMatrixPtr& bias,
+                             MKLDNNMatrixPtr& out) {
+  resetFwdBuffers(in, wgt, bias, out);
 
-  // reset mkldnn forward
-  resetFwd();
-  needResetBwd_ = true;
+  resetFwdPD(fwdPD_, in, wgt, bias, out);
 
-  convertWeightsFromPaddle();
+  resetFwdPipeline(pipeline, fwdPD_, in, wgt, bias, out);
+
+  printValueFormatFlow();
 }
 
-void MKLDNNFcLayer::resetFwd() {
-  bool hasBias = biases_ && biases_->getW();
-  const MatrixPtr& wgt = weight_->getW();
-  const MatrixPtr& bias = hasBias ? biases_->getW() : nullptr;
-  const MatrixPtr& out = output_.value;
+void MKLDNNFcLayer::resetBwd(std::vector<primitive>& pipeline,
+                             MKLDNNMatrixPtr& in,
+                             MKLDNNMatrixPtr& wgt,
+                             MKLDNNMatrixPtr& bias,
+                             MKLDNNMatrixPtr& out) {
+  std::shared_ptr<fc_bwdWgt::primitive_desc> bwdWgtPD;
+  std::shared_ptr<fc_bwdData::primitive_desc> bwdDataPD;
+
+  resetBwdBuffers(in, wgt, bias, out);
+
+  resetBwdWgtPD(bwdWgtPD, wgt, bias, out);
+
+  resetBwdDataPD(bwdDataPD, in, out);
+
+  resetBwdPipeline(pipeline, bwdWgtPD, bwdDataPD, in, wgt, bias, out);
 
+  printGradFormatFlow();
+}
+
+void MKLDNNFcLayer::updateInputData() {
+  inVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
+}
+
+void MKLDNNFcLayer::updateWeights(const UpdateCallback& callback) {
+  weight_->getParameterPtr()->incUpdate(callback);
+  if (biases_ && biases_->getWGrad()) {
+    biases_->getParameterPtr()->incUpdate(callback);
+  }
+}
+
+void MKLDNNFcLayer::resetFwdBuffers(MKLDNNMatrixPtr& in,
+                                    MKLDNNMatrixPtr& wgt,
+                                    MKLDNNMatrixPtr& bias,
+                                    MKLDNNMatrixPtr& out) {
+  resetInValue(in);
+
+  resetWgtBiasValue(wgt, bias);
+
+  resetOutValue(out);
+}
+
+void MKLDNNFcLayer::resetInValue(MKLDNNMatrixPtr& in) {
   if (inputIsOnlyMKLDNN()) {
-    const MatrixPtr& in = getInputValue(0);
-    inVal_ = std::dynamic_pointer_cast<MKLDNNMatrix>(in);
-    CHECK(inVal_) << "Input should be MKLDNNMatrix";
+    const MatrixPtr& dnnIn = getInputValue(0);
+    in = std::dynamic_pointer_cast<MKLDNNMatrix>(dnnIn);
+    CHECK(in) << "Input should be MKLDNNMatrix";
   } else {
     CHECK_EQ(getPrev(0)->getDeviceId(), CPU_DEVICE) << "Only support CPU yet";
-    const MatrixPtr& in = getInputValue(0, CPU_DEVICE);
-    inVal_ = MKLDNNMatrix::create(
-        in, memory::dims{bs_, ic_, ih_, iw_}, format::nchw, engine_);
+    const MatrixPtr& cpuIn = getInputValue(0, CPU_DEVICE);
+    in = MKLDNNMatrix::create(
+        cpuIn, {bs_, ic_, ih_, iw_}, format::nchw, engine_);
   }
-  inVal_->downSpatial();
-  wgtVal_ = MKLDNNMatrix::create(
-      wgt, memory::dims{oc_, ic_, ih_, iw_}, format::oihw, engine_);
-  wgtVal_->downSpatial();
-  biasVal_ =
-      hasBias ? MKLDNNMatrix::create(bias, {oc_}, format::x, engine_) : nullptr;
-  outVal_ = MKLDNNMatrix::create(out, {bs_, oc_}, format::nc, engine_);
+  in->downSpatial();
+}
+
+void MKLDNNFcLayer::resetWgtBiasValue(MKLDNNMatrixPtr& wgt,
+                                      MKLDNNMatrixPtr& bias) {
+  wgt = MKLDNNMatrix::create(
+      weight_->getW(), {oc_, ic_, ih_, iw_}, format::oihw, engine_);
+  wgt->downSpatial();
+
+  bias = (biases_ && biases_->getW())
+             ? MKLDNNMatrix::create(biases_->getW(), {oc_}, format::x, engine_)
+             : nullptr;
+}
 
+void MKLDNNFcLayer::resetOutValue(MKLDNNMatrixPtr& out) {
+  out = MKLDNNMatrix::create(output_.value, {bs_, oc_}, format::nc, engine_);
   // change original output value to mkldnn output value
-  output_.value = std::dynamic_pointer_cast<Matrix>(outVal_);
+  output_.value = std::dynamic_pointer_cast<Matrix>(out);
   if (!outputIsOnlyMKLDNN()) {
-    convertOutputToOtherDevice();
+    // fc cpu output value do not need create convert
+    // just share point
+    getOutput(CPU_DEVICE).value->setData(output_.value->getData());
   }
+}
 
-  // create forward handle
+void MKLDNNFcLayer::resetFwdPD(std::shared_ptr<fc_fwd::primitive_desc>& pd,
+                               MKLDNNMatrixPtr in,
+                               MKLDNNMatrixPtr wgt,
+                               MKLDNNMatrixPtr bias,
+                               MKLDNNMatrixPtr out) {
+  CHECK(in);
+  CHECK(wgt);
+  CHECK(out);
   prop_kind pk = prop_kind::forward;
-  fc_fwd::desc fwdDesc = hasBias ? fc_fwd::desc(pk,
-                                                inVal_->getMemoryDesc(),
-                                                wgtVal_->getMemoryDesc(),
-                                                biasVal_->getMemoryDesc(),
-                                                outVal_->getMemoryDesc())
-                                 : fc_fwd::desc(pk,
-                                                inVal_->getMemoryDesc(),
-                                                wgtVal_->getMemoryDesc(),
-                                                outVal_->getMemoryDesc());
-  fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_);
-  if (hasBias) {
-    fwd_.reset(new fc_fwd(fwdPD, *inVal_, *wgtVal_, *biasVal_, *outVal_));
+  fc_fwd::desc fwdDesc = bias != nullptr ? fc_fwd::desc(pk,
+                                                        in->getMemoryDesc(),
+                                                        wgt->getMemoryDesc(),
+                                                        bias->getMemoryDesc(),
+                                                        out->getMemoryDesc())
+                                         : fc_fwd::desc(pk,
+                                                        in->getMemoryDesc(),
+                                                        wgt->getMemoryDesc(),
+                                                        out->getMemoryDesc());
+  pd.reset(new fc_fwd::primitive_desc(fwdDesc, engine_));
+}
+
+void MKLDNNFcLayer::resetFwdPipeline(
+    std::vector<primitive>& pipeline,
+    std::shared_ptr<fc_fwd::primitive_desc>& pd,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias,
+    MKLDNNMatrixPtr& out) {
+  pipeline.clear();
+
+  if (bias) {
+    fwd_.reset(new fc_fwd(*pd, *in, *wgt, *bias, *out));
   } else {
-    fwd_.reset(new fc_fwd(fwdPD, *inVal_, *wgtVal_, *outVal_));
+    fwd_.reset(new fc_fwd(*pd, *in, *wgt, *out));
   }
-  printValueFormatFlow();
 
-  pipelineFwd_.clear();
-  pipelineFwd_.push_back(*fwd_);
+  pipeline.push_back(*fwd_);
 }
 
-void MKLDNNFcLayer::resetBwd() {
-  if (!needResetBwd_) {
-    return;
-  }
-  needResetBwd_ = false;
-  bool hasBias = biases_ && biases_->getWGrad();
+void MKLDNNFcLayer::resetBwdBuffers(MKLDNNMatrixPtr& in,
+                                    MKLDNNMatrixPtr& wgt,
+                                    MKLDNNMatrixPtr& bias,
+                                    MKLDNNMatrixPtr& out) {
+  resetOutGrad(out);
 
-  /// backward weight
-  CHECK(inVal_) << "Should have input value";
-  const MatrixPtr& wgt = weight_->getWGrad();
-  const MatrixPtr& bias = hasBias ? biases_->getWGrad() : nullptr;
+  resetWgtBiasGrad(wgt, bias);
+
+  resetInGrad(in);
+}
 
+void MKLDNNFcLayer::resetOutGrad(MKLDNNMatrixPtr& out) {
   // TODO(TJ): merge outgrad
   int device = outputIsOnlyMKLDNN() ? MKLDNN_DEVICE : CPU_DEVICE;
   // for MKLDNN device:
@@ -202,101 +241,88 @@ void MKLDNNFcLayer::resetBwd() {
   // for CPU device:
   // fc do not need to convert from cpu device since output is always nc format
   // only need create from cpu device
-  const MatrixPtr& out = getOutput(device).grad;
-  outGrad_ = MKLDNNMatrix::create(out, outVal_->getPrimitiveDesc());
-  wgtGrad_ = MKLDNNMatrix::create(wgt, wgtVal_->getPrimitiveDesc());
-  biasGrad_ = hasBias ? MKLDNNMatrix::create(bias, biasVal_->getPrimitiveDesc())
-                      : nullptr;
-
-  // create memory primitive desc
-  fc_fwd::desc fwdDesc = fc_fwd::desc(prop_kind::forward,
-                                      inVal_->getMemoryDesc(),
-                                      wgtGrad_->getMemoryDesc(),
-                                      outGrad_->getMemoryDesc());
-  fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_);
-  fc_bwdWgt::desc bwdWgtDesc = hasBias
-                                   ? fc_bwdWgt::desc(inVal_->getMemoryDesc(),
-                                                     wgtGrad_->getMemoryDesc(),
-                                                     biasGrad_->getMemoryDesc(),
-                                                     outGrad_->getMemoryDesc())
-                                   : fc_bwdWgt::desc(inVal_->getMemoryDesc(),
-                                                     wgtGrad_->getMemoryDesc(),
-                                                     outGrad_->getMemoryDesc());
-  fc_bwdWgt::primitive_desc bwdWgtPD =
-      fc_bwdWgt::primitive_desc(bwdWgtDesc, engine_, fwdPD);
-
-  if (hasBias) {
-    bwdWgt_.reset(
-        new fc_bwdWgt(bwdWgtPD, *inVal_, *outGrad_, *wgtGrad_, *biasGrad_));
-  } else {
-    bwdWgt_.reset(new fc_bwdWgt(bwdWgtPD, *inVal_, *outGrad_, *wgtGrad_));
-  }
-  pipelineBwd_.clear();
-  pipelineBwd_.push_back(*bwdWgt_);
+  CHECK(outVal_);
+  out =
+      MKLDNNMatrix::create(getOutput(device).grad, outVal_->getPrimitiveDesc());
+}
 
-  /// backward data
-  device = inputIsOnlyMKLDNN() ? MKLDNN_DEVICE : CPU_DEVICE;
-  const MatrixPtr& in = getInputGrad(0, device);
-  if (in == nullptr) {
+void MKLDNNFcLayer::resetWgtBiasGrad(MKLDNNMatrixPtr& wgt,
+                                     MKLDNNMatrixPtr& bias) {
+  CHECK(wgtVal_);
+  wgt = MKLDNNMatrix::create(weight_->getWGrad(), wgtVal_->getPrimitiveDesc());
+
+  bias = nullptr;
+  if (biasVal_ == nullptr) {
     return;
   }
-  if (getInput(0, device).getAllCount() > 1) {
-    // TODO(TJ): use outputMaps_ ways when merge outgrad done
-  } else {
-    inGrad_ = MKLDNNMatrix::create(in, inVal_->getPrimitiveDesc());
-  }
-
-  fc_bwdData::desc bwdDataDesc = fc_bwdData::desc(inVal_->getMemoryDesc(),
-                                                  wgtGrad_->getMemoryDesc(),
-                                                  outGrad_->getMemoryDesc());
-  fc_bwdData::primitive_desc bwdDataPD =
-      fc_bwdData::primitive_desc(bwdDataDesc, engine_, fwdPD);
-
-  CHECK(wgtVal_) << "Should have weight memory";
-  bwdData_.reset(new fc_bwdData(bwdDataPD, *outGrad_, *wgtVal_, *inGrad_));
-  printGradFormatFlow();
-  pipelineBwd_.push_back(*bwdData_);
+  bias =
+      MKLDNNMatrix::create(biases_->getWGrad(), biasVal_->getPrimitiveDesc());
 }
 
-void MKLDNNFcLayer::forward(PassType passType) {
-  Layer::forward(passType);
-  reshape();
-
-  {
-    REGISTER_TIMER_INFO("mkldnn_FwdTimer", getName().c_str());
-    syncInputValue();
-
-    // just submit forward pipeline
-    stream_->submit(pipelineFwd_);
+void MKLDNNFcLayer::resetInGrad(MKLDNNMatrixPtr& in) {
+  in = nullptr;
+  const MatrixPtr& inGrad = inputLayers_[0]->getOutput().grad;
+  if (inGrad == nullptr) {
+    return;
   }
+  // TODO(TJ): use outputMaps_ ways to get the inGrad_ when merge outgrad done
+  CHECK(inVal_);
+  in = MKLDNNMatrix::create(inGrad, inVal_->getPrimitiveDesc());
+}
 
-  /* activation */ {
-    REGISTER_TIMER_INFO("FwActTimer", getName().c_str());
-    forwardActivation();
-  }
+void MKLDNNFcLayer::resetBwdWgtPD(
+    std::shared_ptr<fc_bwdWgt::primitive_desc>& pd,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias,
+    MKLDNNMatrixPtr& out) {
+  CHECK(inVal_);
+  fc_bwdWgt::desc bwdWgtDesc = bias ? fc_bwdWgt::desc(inVal_->getMemoryDesc(),
+                                                      wgt->getMemoryDesc(),
+                                                      bias->getMemoryDesc(),
+                                                      out->getMemoryDesc())
+                                    : fc_bwdWgt::desc(inVal_->getMemoryDesc(),
+                                                      wgt->getMemoryDesc(),
+                                                      out->getMemoryDesc());
+  pd.reset(new fc_bwdWgt::primitive_desc(bwdWgtDesc, engine_, *fwdPD_));
 }
 
-void MKLDNNFcLayer::backward(const UpdateCallback& callback) {
-  /* Do derivation */ {
-    REGISTER_TIMER_INFO("BpActTimer", getName().c_str());
-    backwardActivation();
+void MKLDNNFcLayer::resetBwdDataPD(
+    std::shared_ptr<fc_bwdData::primitive_desc>& pd,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& out) {
+  pd = nullptr;
+  if (in == nullptr) {
+    return;
   }
+  CHECK(wgtVal_);
+  fc_bwdData::desc bwdDataDesc = fc_bwdData::desc(
+      in->getMemoryDesc(), wgtVal_->getMemoryDesc(), out->getMemoryDesc());
+  pd.reset(new fc_bwdData::primitive_desc(bwdDataDesc, engine_, *fwdPD_));
+}
 
-  {
-    REGISTER_TIMER_INFO("mkldnn_bwdTimer", getName().c_str());
-    resetBwd();
-
-    syncOutputGrad();
-    // just sumbmit backward pipeline
-    stream_->submit(pipelineBwd_);
+void MKLDNNFcLayer::resetBwdPipeline(
+    std::vector<primitive>& pipeline,
+    std::shared_ptr<fc_bwdWgt::primitive_desc>& bwdWgtPD,
+    std::shared_ptr<fc_bwdData::primitive_desc>& bwdDataPD,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias,
+    MKLDNNMatrixPtr& out) {
+  pipeline.clear();
+  CHECK(inVal_);
+  if (bias) {
+    bwdWgt_.reset(new fc_bwdWgt(*bwdWgtPD, *inVal_, *out, *wgt, *bias));
+  } else {
+    bwdWgt_.reset(new fc_bwdWgt(*bwdWgtPD, *inVal_, *out, *wgt));
   }
+  pipeline.push_back(*bwdWgt_);
 
-  {
-    REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
-    weight_->getParameterPtr()->incUpdate(callback);
-    if (biases_ && biases_->getWGrad()) {
-      biases_->getParameterPtr()->incUpdate(callback);
-    }
+  if (bwdDataPD == nullptr) {
+    return;
   }
+  CHECK(wgtVal_) << "Should have weight memory";
+  bwdData_.reset(new fc_bwdData(*bwdDataPD, *out, *wgtVal_, *in));
+  pipeline.push_back(*bwdData_);
 }
+
 }  // namespace paddle
diff --git a/paddle/gserver/layers/MKLDNNFcLayer.h b/paddle/gserver/layers/MKLDNNFcLayer.h
index e138a6faf1..c76878aafa 100644
--- a/paddle/gserver/layers/MKLDNNFcLayer.h
+++ b/paddle/gserver/layers/MKLDNNFcLayer.h
@@ -18,6 +18,9 @@ limitations under the License. */
 #include "mkldnn.hpp"
 
 namespace paddle {
+typedef mkldnn::inner_product_forward fc_fwd;
+typedef mkldnn::inner_product_backward_weights fc_bwdWgt;
+typedef mkldnn::inner_product_backward_data fc_bwdData;
 
 /**
  * @brief A subclass of MKLDNNLayer fc layer.
@@ -32,6 +35,9 @@ protected:
   // if has already init the weight
   bool hasInitedWgt_;
 
+  // save forward primitive_desc, which can be used backward
+  std::shared_ptr<fc_fwd::primitive_desc> fwdPD_;
+
   // fc weight and bias
   std::unique_ptr<Weight> weight_;
   std::unique_ptr<Weight> biases_;
@@ -45,35 +51,81 @@ public:
   bool init(const LayerMap& layerMap,
             const ParameterMap& parameterMap) override;
 
-  void convertWeightsFromPaddle() override;
+  void reshape(
+      int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) override;
 
-  void convertWeightsToPaddle() override;
+  void resetFwd(std::vector<mkldnn::primitive>& pipeline,
+                MKLDNNMatrixPtr& in,
+                MKLDNNMatrixPtr& wgt,
+                MKLDNNMatrixPtr& bias,
+                MKLDNNMatrixPtr& out) override;
 
-  void forward(PassType passType) override;
+  void resetBwd(std::vector<mkldnn::primitive>& pipeline,
+                MKLDNNMatrixPtr& in,
+                MKLDNNMatrixPtr& wgt,
+                MKLDNNMatrixPtr& bias,
+                MKLDNNMatrixPtr& out) override;
 
-  void backward(const UpdateCallback& callback) override;
+  void updateInputData() override;
 
-protected:
-  /**
-   * reshape the input image sizes
-   * and reset output buffer size
-   * and reset mkldnn forward
-   */
-  void reshape();
+  void updateWeights(const UpdateCallback& callback) override;
 
+  void convertWeightsFromPaddle() override;
+
+  void convertWeightsToPaddle() override;
+
+protected:
   /**
-   * reset the forward primitve and memory
-   * only would be called when input size changes
+   * Forward functions: reset buffers(input, output, weight and bias),
+   *                    reset primitive descriptor,
+   *                    reset pipeline.
    */
-  void resetFwd();
+  void resetFwdBuffers(MKLDNNMatrixPtr& in,
+                       MKLDNNMatrixPtr& wgt,
+                       MKLDNNMatrixPtr& bias,
+                       MKLDNNMatrixPtr& out);
+  void resetInValue(MKLDNNMatrixPtr& in);
+  void resetWgtBiasValue(MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias);
+  void resetOutValue(MKLDNNMatrixPtr& out);
+  void resetFwdPD(std::shared_ptr<fc_fwd::primitive_desc>& pd,
+                  MKLDNNMatrixPtr in,
+                  MKLDNNMatrixPtr wgt,
+                  MKLDNNMatrixPtr bias,
+                  MKLDNNMatrixPtr out);
+  void resetFwdPipeline(std::vector<mkldnn::primitive>& pipeline,
+                        std::shared_ptr<fc_fwd::primitive_desc>& pd,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& wgt,
+                        MKLDNNMatrixPtr& bias,
+                        MKLDNNMatrixPtr& out);
 
   /**
-   * reset the backward primitve and memory for mkldnn fc
-   * only would be called when needed
+   * Backward functions: reset buffers(input, output, weight and bias),
+   *                     reset primitive descriptor for backward weight,
+   *                     reset primitive descriptor for backward data,
+   *                     reset pipeline.
    */
-  void resetBwd();
-
-  void convertOutputToOtherDevice() override;
+  void resetBwdBuffers(MKLDNNMatrixPtr& in,
+                       MKLDNNMatrixPtr& wgt,
+                       MKLDNNMatrixPtr& bias,
+                       MKLDNNMatrixPtr& out);
+  void resetOutGrad(MKLDNNMatrixPtr& out);
+  void resetWgtBiasGrad(MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias);
+  void resetInGrad(MKLDNNMatrixPtr& in);
+  void resetBwdWgtPD(std::shared_ptr<fc_bwdWgt::primitive_desc>& pd,
+                     MKLDNNMatrixPtr& wgt,
+                     MKLDNNMatrixPtr& bias,
+                     MKLDNNMatrixPtr& out);
+  void resetBwdDataPD(std::shared_ptr<fc_bwdData::primitive_desc>& pd,
+                      MKLDNNMatrixPtr& in,
+                      MKLDNNMatrixPtr& out);
+  void resetBwdPipeline(std::vector<mkldnn::primitive>& pipeline,
+                        std::shared_ptr<fc_bwdWgt::primitive_desc>& bwdWgtPD,
+                        std::shared_ptr<fc_bwdData::primitive_desc>& bwdDataPD,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& wgt,
+                        MKLDNNMatrixPtr& bias,
+                        MKLDNNMatrixPtr& out);
 };
 
 }  // namespace paddle
diff --git a/paddle/gserver/layers/MKLDNNLayer.h b/paddle/gserver/layers/MKLDNNLayer.h
index b983b833d5..169679c829 100644
--- a/paddle/gserver/layers/MKLDNNLayer.h
+++ b/paddle/gserver/layers/MKLDNNLayer.h
@@ -19,6 +19,7 @@ limitations under the License. */
 #include "MKLDNNBase.h"
 #include "mkldnn.hpp"
 #include "paddle/math/MKLDNNMatrix.h"
+#include "paddle/utils/Stat.h"
 
 DECLARE_bool(use_mkldnn);
 
@@ -33,6 +34,8 @@ typedef std::shared_ptr<MKLDNNLayer> MKLDNNLayerPtr;
  */
 class MKLDNNLayer : public Layer {
 protected:
+  // input value element count
+  size_t inputElemenCnt_;
   // batch size
   int bs_;
   // input image channel, height and width
@@ -52,7 +55,7 @@ protected:
   std::vector<mkldnn::primitive> pipelineFwd_;
   std::vector<mkldnn::primitive> pipelineBwd_;
 
-  // MKLDNNMatrixPtr
+  // MKLDNNMatrixPtr with internal format
   MKLDNNMatrixPtr inVal_;
   MKLDNNMatrixPtr inGrad_;
   MKLDNNMatrixPtr outVal_;
@@ -65,6 +68,7 @@ protected:
 public:
   explicit MKLDNNLayer(const LayerConfig& config)
       : Layer(config),
+        inputElemenCnt_(0),
         bs_(0),
         ic_(0),
         ih_(0),
@@ -95,12 +99,104 @@ public:
     if (!Layer::init(layerMap, parameterMap)) {
       return false;
     }
+    checkCPUOutputsNumber();
 
     stream_.reset(new MKLDNNStream());
     engine_ = CPUEngine::Instance().getEngine();
     return true;
   }
 
+  void forward(PassType passType) override {
+    passType_ = passType;
+
+    {
+      REGISTER_TIMER_INFO("mkldnn_FwdTimer", getName().c_str());
+      CHECK(!inputLayers_.empty());
+      copySeqInfoToOutputs();
+      size_t elemenCnt = inputLayers_[0]->getOutput().value->getElementCnt();
+      if (inputElemenCnt_ != elemenCnt) {
+        // reset when input total sizes changed, not only the batchsize
+        inputElemenCnt_ = elemenCnt;
+        reshape(bs_, ic_, ih_, iw_, oc_, oh_, ow_);
+        resetFwd(pipelineFwd_, inVal_, wgtVal_, biasVal_, outVal_);
+        convertWeightsFromPaddle();
+        needResetBwd_ = true;
+      }
+
+      if (inputLayers_[0]->getType() == "data") {
+        updateInputData();
+      }
+
+      stream_->submit(pipelineFwd_);
+    }
+
+    /* activation */ {
+      REGISTER_TIMER_INFO("FwActTimer", getName().c_str());
+      forwardActivation();
+    }
+  }
+
+  void backward(const UpdateCallback& callback) override {
+    /* Do derivation */ {
+      REGISTER_TIMER_INFO("BpActTimer", getName().c_str());
+      backwardActivation();
+    }
+
+    {
+      REGISTER_TIMER_INFO("mkldnn_bwdTimer", getName().c_str());
+      if (needResetBwd_) {
+        resetBwd(pipelineBwd_, inGrad_, wgtGrad_, biasGrad_, outGrad_);
+        needResetBwd_ = false;
+      }
+
+      stream_->submit(pipelineBwd_);
+    }
+
+    {
+      REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
+      updateWeights(callback);
+    }
+  }
+
+  /**
+   * reshape the input image sizes
+   * and reset output image and buffer size
+   * output channel can not be changed
+   */
+  virtual void reshape(
+      int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) = 0;
+
+  /**
+   * reset the mkldnn forward primitve and memory
+   * only would be called when input size changes
+   */
+  virtual void resetFwd(std::vector<mkldnn::primitive>& pipeline,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& wgt,
+                        MKLDNNMatrixPtr& bias,
+                        MKLDNNMatrixPtr& out) = 0;
+
+  /**
+   * reset the mkldnn backward primitve and memory for mkldnn fc
+   * only would be called when needed
+   */
+  virtual void resetBwd(std::vector<mkldnn::primitive>& pipeline,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& wgt,
+                        MKLDNNMatrixPtr& bias,
+                        MKLDNNMatrixPtr& out) = 0;
+
+  /**
+   * Update input value data when input layer is "data" type.
+   * Since the input value data address might be changed.
+   */
+  virtual void updateInputData() {}
+
+  /**
+   * Update weights and biases if necessary.
+   */
+  virtual void updateWeights(const UpdateCallback& callback) {}
+
   /**
    * convert weight from paddle format to mkldnn format
    * weight_ will be override
@@ -114,10 +210,38 @@ public:
   virtual void convertWeightsToPaddle() {}
 
   /**
-   * convert MKLDNN output to other device.
-   * only support CPU device yet
+   * add this interface as public for unit test
+   */
+  void addOutputArgument(int deviceId) { Layer::addOutputArgument(deviceId); }
+
+protected:
+  /**
+   * reshape the input image sizes and input batchsize
    */
-  virtual void convertOutputToOtherDevice() {}
+  virtual void reshapeInput(int& batchsize, int& height, int& width) {
+    const Argument& input = inputLayers_[0]->getOutput();
+    batchsize = input.getBatchSize();
+    int h = input.getFrameHeight();
+    int w = input.getFrameWidth();
+    if (h != 0) {
+      height = h;
+    }
+    if (w != 0) {
+      width = w;
+    }
+  }
+
+  /**
+   * reshape output image sizes
+   */
+  virtual void reshapeOutput(size_t height, size_t width) {
+    output_.setFrameHeight(height);
+    output_.setFrameWidth(width);
+    for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
+      outputOtherDevice_[i].setFrameHeight(height);
+      outputOtherDevice_[i].setFrameWidth(width);
+    }
+  }
 
   /**
    * print info about sizes
@@ -133,8 +257,8 @@ public:
    */
   virtual void printValueFormatFlow() {
     if (inVal_ && outVal_) {
-      VLOG(MKLDNN_FMTS) << "value format flow --- " << inVal_->getFormat()
-                        << " >>> " << outVal_->getFormat();
+      VLOG(MKLDNN_FMTS) << inVal_->getFormat() << " >>> "
+                        << outVal_->getFormat();
     }
   }
 
@@ -143,29 +267,12 @@ public:
    */
   virtual void printGradFormatFlow() {
     if (inGrad_ && outGrad_) {
-      VLOG(MKLDNN_FMTS) << "grad format flow --- " << inGrad_->getFormat()
-                        << " <<< " << outGrad_->getFormat();
+      VLOG(MKLDNN_FMTS) << inGrad_->getFormat() << " <<< "
+                        << outGrad_->getFormat();
     }
   }
 
 protected:
-  /**
-   * copy image size and sequence info to other device
-   * @note: can not directly use Layer::copyOutputToOtherDevice since here only
-   *        copy base info and do not copy data value
-   */
-  void copyOutputInfoToOtherDevice() {
-    for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
-      outputOtherDevice_[i].setFrameHeight(output_.getFrameHeight());
-      outputOtherDevice_[i].setFrameWidth(output_.getFrameWidth());
-      outputOtherDevice_[i].sequenceStartPositions =
-          output_.sequenceStartPositions;
-      outputOtherDevice_[i].subSequenceStartPositions =
-          output_.subSequenceStartPositions;
-      outputOtherDevice_[i].cpuSequenceDims = output_.cpuSequenceDims;
-    }
-  }
-
   /**
    * If input only has MKLDNN device.
    * Otherwise, only support the previous layer using CPU device.
@@ -193,37 +300,12 @@ protected:
     return outputOtherDevice_.size() == 0;
   }
 
-  /**
-   * Sync input value data
-   */
-  void syncInputValue() {
-    if (inputIsOnlyMKLDNN()) {
-      return;
-    }
-    real* iData = getInputValue(0, CPU_DEVICE)->getData();
-    // update input data
-    // since it might be changed if this is after data layer
-    inVal_->updateData(iData);
-  }
-
-  /**
-   * Sync output grad data
-   */
-  void syncOutputGrad() {
-    if (outputIsOnlyMKLDNN()) {
-      return;
-    }
-
-    // update diff
-    real* oDiff = getOutput(CPU_DEVICE).grad->getData();
-    outGrad_->updateData(oDiff);
-  }
-
   /**
    * Set deviceId of this layer.
    */
   void setDevice(int id) { deviceId_ = id; }
 
+private:
   /**
    * Set deviceId of the params used in this layer.
    */
@@ -247,6 +329,42 @@ protected:
       parameter->setDevice(id);
     }
   }
+
+  /**
+   * Check the cpu device number of outputOtherDevice_.
+   * should have only one at most.
+   */
+  void checkCPUOutputsNumber(int max = 1) {
+    int cnt = 0;
+    for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
+      if (outputOtherDevice_[i].deviceId == CPU_DEVICE) {
+        ++cnt;
+      }
+    }
+    CHECK_LE(cnt, max) << "too much CPU devies";
+  }
+
+  /**
+   * copy SeqInfo from input layer to this output and other output devices.
+   * @note: do not use getInput(0) since it used this deviceId_,
+   *        use "inputLayers_[0]->getOutput()" instead.
+   */
+  void copySeqInfoToOutputs() {
+    if (inputLayers_.empty() || !needSequenceInfo_) {
+      return;
+    }
+    const Argument& input = inputLayers_[0]->getOutput();
+    output_.sequenceStartPositions = input.sequenceStartPositions;
+    output_.subSequenceStartPositions = input.subSequenceStartPositions;
+    output_.cpuSequenceDims = input.cpuSequenceDims;
+    for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
+      outputOtherDevice_[i].sequenceStartPositions =
+          output_.sequenceStartPositions;
+      outputOtherDevice_[i].subSequenceStartPositions =
+          output_.subSequenceStartPositions;
+      outputOtherDevice_[i].cpuSequenceDims = output_.cpuSequenceDims;
+    }
+  }
 };
 
 }  // namespace paddle
diff --git a/paddle/gserver/layers/MKLDNNPoolLayer.cpp b/paddle/gserver/layers/MKLDNNPoolLayer.cpp
new file mode 100644
index 0000000000..48b2f5a4cb
--- /dev/null
+++ b/paddle/gserver/layers/MKLDNNPoolLayer.cpp
@@ -0,0 +1,277 @@
+/* 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 "MKLDNNPoolLayer.h"
+#include "paddle/math/MathUtils.h"
+#include "paddle/utils/Logging.h"
+
+using namespace mkldnn;  // NOLINT
+typedef memory::format format;
+
+namespace paddle {
+
+REGISTER_LAYER(mkldnn_pool, MKLDNNPoolLayer);
+
+bool MKLDNNPoolLayer::init(const LayerMap& layerMap,
+                           const ParameterMap& parameterMap) {
+  if (!MKLDNNLayer::init(layerMap, parameterMap)) {
+    return false;
+  }
+
+  /* the size of inputs for pool-layer is 1 */
+  CHECK_EQ(config_.inputs_size(), 1);
+  const PoolConfig& conf = config_.inputs(0).pool_conf();
+  ic_ = conf.channels();
+  ih_ = conf.img_size_y();
+  iw_ = conf.img_size();
+  oc_ = ic_;
+  oh_ = conf.output_y();
+  ow_ = conf.output_x();
+  fh_ = conf.size_y();
+  fw_ = conf.size_x();
+  ph_ = conf.padding_y();
+  pw_ = conf.padding();
+  sh_ = conf.stride_y();
+  sw_ = conf.stride();
+
+  const std::string& type = conf.pool_type();
+  if (type == "max-projection") {
+    poolAlgo_ = algorithm::pooling_max;
+  } else if (type == "avg-projection") {
+    // paddle only use exclude_padding
+    poolAlgo_ = algorithm::pooling_avg_exclude_padding;
+  } else {
+    LOG(FATAL) << "unknow pooling type!";
+  }
+  return true;
+}
+
+void MKLDNNPoolLayer::reshape(
+    int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) {
+  reshapeInput(bs, ih, iw);
+  // ic_ and oc can not be changed
+  CHECK_EQ(inputElemenCnt_ / bs / ih / iw, (size_t)ic)
+      << "Input channel can not be changed";
+
+  // cal output sizes
+  // paddle used false caffeMode for pooling
+  oh = outputSize(ih, fh_, ph_, sh_, false);
+  ow = outputSize(iw, fw_, pw_, sw_, false);
+  reshapeOutput(oh, ow);
+
+  resizeOutput(bs, oc * oh * ow);
+
+  printSizeInfo();
+}
+
+void MKLDNNPoolLayer::resetFwd(std::vector<primitive>& pipeline,
+                               MKLDNNMatrixPtr& in,
+                               MKLDNNMatrixPtr& wgt,
+                               MKLDNNMatrixPtr& bias,
+                               MKLDNNMatrixPtr& out) {
+  resetFwdBuffers(in, out);
+
+  resetFwdPD(fwdPD_, in, out);
+
+  resetFwdPipeline(pipeline, fwdPD_, in, out);
+
+  printValueFormatFlow();
+}
+
+void MKLDNNPoolLayer::resetBwd(std::vector<primitive>& pipeline,
+                               MKLDNNMatrixPtr& in,
+                               MKLDNNMatrixPtr& wgt,
+                               MKLDNNMatrixPtr& bias,
+                               MKLDNNMatrixPtr& out) {
+  std::shared_ptr<pool_bwd::primitive_desc> pd;
+
+  resetBwdBuffers(in, out);
+
+  resetBwdPD(pd, in, out);
+
+  resetBwdPipeline(pipeline, pd, in, out);
+
+  printGradFormatFlow();
+}
+
+void MKLDNNPoolLayer::updateInputData() {
+  inVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
+}
+
+void MKLDNNPoolLayer::resetFwdBuffers(MKLDNNMatrixPtr& in,
+                                      MKLDNNMatrixPtr& out) {
+  resetInValue(in);
+
+  resetOutValue(out);
+}
+
+void MKLDNNPoolLayer::resetInValue(MKLDNNMatrixPtr& in) {
+  if (inputIsOnlyMKLDNN()) {
+    const MatrixPtr& dnnIn = getInputValue(0);
+    in = std::dynamic_pointer_cast<MKLDNNMatrix>(dnnIn);
+    CHECK(in) << "Input should be MKLDNNMatrix";
+  } else {
+    CHECK_EQ(getPrev(0)->getDeviceId(), CPU_DEVICE) << "Only support CPU yet";
+    const MatrixPtr& cpuIn = getInputValue(0, CPU_DEVICE);
+    in = MKLDNNMatrix::create(
+        cpuIn, {bs_, ic_, ih_, iw_}, format::nchw, engine_);
+  }
+}
+
+void MKLDNNPoolLayer::resetOutValue(MKLDNNMatrixPtr& out) {
+  CHECK(inVal_) << "Should reset input value first";
+  memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
+  out = MKLDNNMatrix::create(
+      output_.value, outDims, inVal_->getFormat(), engine_);
+  output_.value = std::dynamic_pointer_cast<Matrix>(out);
+
+  // create reorder if output value has cpu device and pd do not match
+  cpuOutVal_ = nullptr;
+  cvtOutVal_ = nullptr;
+  if (!outputIsOnlyMKLDNN()) {
+    const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).value;
+    cpuOutVal_ = MKLDNNMatrix::create(cpuOut, outDims, format::nchw, engine_);
+    if (cpuOutVal_->getPrimitiveDesc() != out->getPrimitiveDesc()) {
+      cvtOutVal_ = MKLDNNMatrix::createReorder(out, cpuOutVal_);
+      CHECK(cvtOutVal_) << "should not be emptry";
+    } else {
+      // CPU output share the same data of MKLDNN output
+      cpuOut->setData(out->getData());
+      cpuOutVal_ = out;
+    }
+  }
+}
+
+void MKLDNNPoolLayer::resetFwdPD(std::shared_ptr<pool_fwd::primitive_desc>& pd,
+                                 MKLDNNMatrixPtr in,
+                                 MKLDNNMatrixPtr out) {
+  memory::dims inDims = memory::dims{bs_, ic_, ih_, iw_};
+  memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
+  memory::dims kernels = memory::dims{fh_, fw_};
+  memory::dims strides = memory::dims{sh_, sw_};
+  memory::dims padL = memory::dims{ph_, pw_};
+  memory::dims padR = getPaddingR();
+  padding_kind padKind = padding_kind::zero;
+  prop_kind pk = passType_ == PASS_TEST ? prop_kind::forward_scoring
+                                        : prop_kind::forward_training;
+  auto fwdDesc = pool_fwd::desc(pk,
+                                poolAlgo_,
+                                in->getMemoryDesc(),
+                                out->getMemoryDesc(),
+                                strides,
+                                kernels,
+                                padL,
+                                padR,
+                                padKind);
+  pd.reset(new pool_fwd::primitive_desc(fwdDesc, engine_));
+
+  // prepare workspace if necessary
+  workspace_ =
+      (passType_ != PASS_TEST && poolAlgo_ == algorithm::pooling_max)
+          ? std::make_shared<memory>(memory(pd->workspace_primitive_desc()))
+          : nullptr;
+}
+
+void MKLDNNPoolLayer::resetFwdPipeline(
+    std::vector<primitive>& pipeline,
+    std::shared_ptr<pool_fwd::primitive_desc>& pd,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& out) {
+  pipeline.clear();
+  fwd_ = workspace_
+             ? std::make_shared<pool_fwd>(pool_fwd(*pd, *in, *out, *workspace_))
+             : std::make_shared<pool_fwd>(pool_fwd(*pd, *in, *out));
+  pipeline.push_back(*fwd_);
+
+  if (cvtOutVal_) {
+    pipeline.push_back(*cvtOutVal_);
+  }
+}
+
+void MKLDNNPoolLayer::resetBwdBuffers(MKLDNNMatrixPtr& in,
+                                      MKLDNNMatrixPtr& out) {
+  resetOutGrad(out);
+
+  resetInGrad(in);
+}
+void MKLDNNPoolLayer::resetOutGrad(MKLDNNMatrixPtr& out) {
+  CHECK(outVal_) << "Should have output value";
+  out = MKLDNNMatrix::create(output_.grad, outVal_->getPrimitiveDesc());
+
+  // create reorder if output value has cpu device and pd do not match
+  cpuOutGrad_ = nullptr;
+  cvtOutGrad_ = nullptr;
+  if (!outputIsOnlyMKLDNN()) {
+    const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).grad;
+    cpuOutGrad_ = MKLDNNMatrix::create(
+        cpuOut, memory::dims{bs_, oc_, oh_, ow_}, format::nchw, engine_);
+    if (cpuOutGrad_->getPrimitiveDesc() != out->getPrimitiveDesc()) {
+      cvtOutGrad_ = MKLDNNMatrix::createReorder(cpuOutGrad_, out);
+      CHECK(cvtOutGrad_) << "should not be emptry";
+    } else {
+      // share the same data of CPU output
+      output_.grad->setData(cpuOut->getData());
+      out = cpuOutGrad_;
+    }
+  }
+}
+
+void MKLDNNPoolLayer::resetInGrad(MKLDNNMatrixPtr& in) {
+  in = nullptr;
+  const MatrixPtr& inGrad = inputLayers_[0]->getOutput().grad;
+  if (inGrad == nullptr) {
+    return;
+  }
+  CHECK(inVal_);
+  in = MKLDNNMatrix::create(inGrad, inVal_->getPrimitiveDesc());
+}
+
+void MKLDNNPoolLayer::resetBwdPD(std::shared_ptr<pool_bwd::primitive_desc>& pd,
+                                 MKLDNNMatrixPtr& in,
+                                 MKLDNNMatrixPtr& out) {
+  memory::dims kernels = memory::dims{fh_, fw_};
+  memory::dims strides = memory::dims{sh_, sw_};
+  memory::dims padL = memory::dims{ph_, pw_};
+  memory::dims padR = getPaddingR();
+  CHECK(in);
+  CHECK(out);
+  auto bwdDesc = pool_bwd::desc(poolAlgo_,
+                                in->getMemoryDesc(),
+                                out->getMemoryDesc(),
+                                strides,
+                                kernels,
+                                padL,
+                                padR,
+                                padding_kind::zero);
+  pd.reset(new pool_bwd::primitive_desc(bwdDesc, engine_, *fwdPD_));
+}
+
+void MKLDNNPoolLayer::resetBwdPipeline(
+    std::vector<primitive>& pipeline,
+    std::shared_ptr<pool_bwd::primitive_desc>& pd,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& out) {
+  pipeline.clear();
+  if (cvtOutGrad_) {
+    pipeline.push_back(*cvtOutGrad_);
+  }
+
+  bwdData_ =
+      workspace_
+          ? std::make_shared<pool_bwd>(pool_bwd(*pd, *out, *workspace_, *in))
+          : std::make_shared<pool_bwd>(pool_bwd(*pd, *out, *in));
+  pipeline.push_back(*bwdData_);
+}
+
+}  // namespace paddle
diff --git a/paddle/gserver/layers/MKLDNNPoolLayer.h b/paddle/gserver/layers/MKLDNNPoolLayer.h
new file mode 100644
index 0000000000..891e15a7ef
--- /dev/null
+++ b/paddle/gserver/layers/MKLDNNPoolLayer.h
@@ -0,0 +1,138 @@
+/* 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 {
+typedef mkldnn::pooling_forward pool_fwd;
+typedef mkldnn::pooling_backward pool_bwd;
+
+/**
+ * @brief A subclass of MKLDNNLayer pool layer.
+ *
+ * The config file api is mkldnn_pool
+ */
+class MKLDNNPoolLayer : public MKLDNNLayer {
+protected:
+  // padding height and width
+  int ph_, pw_;
+  // stride height and width
+  int sh_, sw_;
+  // filter(kenerl) height and width
+  int fh_, fw_;
+
+  // pooling_avg or pooling_max
+  mkldnn::algorithm poolAlgo_;
+
+  // MKLDNNMatrixPtr which should be created from CPU Device
+  MKLDNNMatrixPtr cpuOutVal_;
+  MKLDNNMatrixPtr cpuOutGrad_;
+  // convert handle between CPU device and MKLDNN device
+  std::shared_ptr<mkldnn::reorder> cvtOutVal_;
+  std::shared_ptr<mkldnn::reorder> cvtOutGrad_;
+
+  // save forward primitive_desc, which can be used backward
+  std::shared_ptr<pool_fwd::primitive_desc> fwdPD_;
+  // according to https://github.com/01org/mkl-dnn/blob/master/tests/gtests/
+  // test_pooling_forward.cpp, pool need workspace for backward
+  std::shared_ptr<mkldnn::memory> workspace_;
+
+public:
+  explicit MKLDNNPoolLayer(const LayerConfig& config) : MKLDNNLayer(config) {}
+
+  ~MKLDNNPoolLayer() {}
+
+  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 updateInputData() override;
+
+  void printSizeInfo() override {
+    MKLDNNLayer::printSizeInfo();
+    VLOG(MKLDNN_SIZES) << getName() << ": fh: " << fh_ << ", fw: " << fw_
+                       << ": ph: " << ph_ << ", pw: " << pw_ << ", sh: " << sh_
+                       << ", sw: " << sw_;
+  }
+
+protected:
+  /**
+   * Forward functions: reset buffers(input, output),
+   *                    reset primitive descriptor,
+   *                    reset pipeline.
+   */
+  void resetFwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& out);
+  void resetInValue(MKLDNNMatrixPtr& in);
+  void resetOutValue(MKLDNNMatrixPtr& out);
+  void resetFwdPD(std::shared_ptr<pool_fwd::primitive_desc>& pd,
+                  MKLDNNMatrixPtr in,
+                  MKLDNNMatrixPtr out);
+  void resetFwdPipeline(std::vector<mkldnn::primitive>& pipeline,
+                        std::shared_ptr<pool_fwd::primitive_desc>& pd,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& out);
+
+  /**
+   * Backward functions: reset buffers(input, output),
+   *                     reset primitive descriptor,
+   *                     reset pipeline.
+   */
+  void resetBwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& out);
+  void resetOutGrad(MKLDNNMatrixPtr& out);
+  void resetInGrad(MKLDNNMatrixPtr& in);
+  void resetBwdPD(std::shared_ptr<pool_bwd::primitive_desc>& pd,
+                  MKLDNNMatrixPtr& in,
+                  MKLDNNMatrixPtr& out);
+  void resetBwdPipeline(std::vector<mkldnn::primitive>& pipeline,
+                        std::shared_ptr<pool_bwd::primitive_desc>& pd,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& out);
+
+  /**
+   * get padding_r according to
+   * https://github.com/01org/mkl-dnn/blob/master/tests/gtests/
+   * test_pooling_forward.cpp
+   */
+  mkldnn::memory::dims getPaddingR() const {
+    mkldnn::memory::dims padR = {ph_, pw_};
+    for (int i = 0; i < 2; ++i) {
+      if ((ih_ + ph_ + padR[0] - fh_) / sh_ + 1 < oh_) {
+        ++padR[0];
+      }
+      if ((iw_ + pw_ + padR[1] - fw_) / sw_ + 1 < ow_) {
+        ++padR[1];
+      }
+    }
+    return padR;
+  }
+};
+
+}  // namespace paddle
diff --git a/paddle/gserver/layers/SwitchOrderLayer.cpp b/paddle/gserver/layers/SwitchOrderLayer.cpp
index d7eee6eaf0..e97809141a 100644
--- a/paddle/gserver/layers/SwitchOrderLayer.cpp
+++ b/paddle/gserver/layers/SwitchOrderLayer.cpp
@@ -83,8 +83,7 @@ void SwitchOrderLayer::forward(PassType passType) {
   setOutDims();
   resetOutput(outDims_[0], outDims_[1] * outDims_[2] * outDims_[3]);
   if (heightAxis_.size() > 0) {
-    getOutputValue()->reshape(reshapeHeight_, reshapeWidth_);
-    getOutputGrad()->reshape(reshapeHeight_, reshapeWidth_);
+    resetOutput(reshapeHeight_, reshapeWidth_);
   }
 
   // switch NCHW to NHWC
diff --git a/paddle/gserver/tests/MKLDNNTester.cpp b/paddle/gserver/tests/MKLDNNTester.cpp
index de1635be2a..2f48e5b2d3 100644
--- a/paddle/gserver/tests/MKLDNNTester.cpp
+++ b/paddle/gserver/tests/MKLDNNTester.cpp
@@ -63,8 +63,12 @@ void MKLDNNTester::reset(const TestConfig& dnn,
     initTestLayer(
         configs_[i], &(layerMaps_[i]), &(parameters_[i]), &(testLayers_[i]));
   }
-  dnnLayer_ = testLayers_[DNN];
   refLayer_ = testLayers_[REF];
+  dnnLayer_ = std::dynamic_pointer_cast<MKLDNNLayer>(testLayers_[DNN]);
+  CHECK(dnnLayer_);
+  // for comparison with Paddle reference results,
+  // need manually add cpu device output for test
+  dnnLayer_->addOutputArgument(CPU_DEVICE);
   EXPECT_EQ(dataLayers_[DNN].size(), dataLayers_[REF].size());
   EXPECT_EQ(parameters_[DNN].size(), parameters_[REF].size());
 
@@ -109,20 +113,22 @@ void MKLDNNTester::randomBotDatas() {
 
 void MKLDNNTester::randomTopDiffs() {
   refLayer_->getOutputGrad()->randomizeUniform();
-  dnnLayer_->getOutputGrad()->copyFrom(*(refLayer_->getOutputGrad()));
-  VLOG(lvl_) << "Random dom Backward Input, TopDiff: ";
+  dnnLayer_->getOutput(CPU_DEVICE)
+      .grad->copyFrom(*(refLayer_->getOutputGrad()));
+  VLOG(lvl_) << "Random Backward Input, TopDiff: ";
   printMatrix(refLayer_->getOutputGrad());
 }
 
 void MKLDNNTester::checkForward() {
-  printTopDatas();
-  double delta = compareMatrix(testLayers_[DNN]->getOutputValue(),
-                               testLayers_[REF]->getOutputValue());
   VLOG(MKLDNN_ALL) << "Check Forward";
+  printTopDatas();
+  double delta = compareMatrix(dnnLayer_->getOutput(-1).value,
+                               refLayer_->getOutputValue());
   EXPECT_LE(fabs(delta), eps_);
 }
 
 void MKLDNNTester::checkBackwardData() {
+  VLOG(MKLDNN_ALL) << "Check Backward Data";
   // TODO(TJ): uncomment me when batch norm ready
   // const bool isBN = dnnLayer_->getType() == "mkldnn_batch_norm";
   for (size_t i = 0; i < dataLayers_[DNN].size(); ++i) {
@@ -144,14 +150,12 @@ void MKLDNNTester::checkBackwardData() {
 }
 
 void MKLDNNTester::checkBackwardWgts() {
+  VLOG(MKLDNN_ALL) << "Check Backward Weight";
   CHECK_EQ(parameters_[DNN].size(), parameters_[REF].size());
   vector<VectorPtr> dnnWgts;  // used to temply save mkldnn weights
   saveWgt(parameters_[DNN], dnnWgts);
 
-  const MKLDNNLayerPtr dnnlayer =
-      std::dynamic_pointer_cast<MKLDNNLayer>(dnnLayer_);
-  CHECK(dnnlayer);
-  dnnlayer->convertWeightsToPaddle();
+  dnnLayer_->convertWeightsToPaddle();
   for (size_t i = 0; i < parameters_[DNN].size(); ++i) {
     const VectorPtr& dnn = parameters_[DNN][i]->getBuf(PARAMETER_VALUE);
     const VectorPtr& ref = parameters_[REF][i]->getBuf(PARAMETER_VALUE);
@@ -189,38 +193,38 @@ void MKLDNNTester::restoreWgt(const vector<VectorPtr>& from,
 }
 
 // clear parameters grad
-void MKLDNNTester::clearWgtDiffs() {
+void MKLDNNTester::clearWgtDiffs(size_t id) {
+  CHECK_LE(id, parameters_.size());
   for (size_t n = 0; n < parameters_.size(); ++n) {
-    for (size_t i = 0; i < parameters_[n].size(); ++i) {
-      const VectorPtr& grad = parameters_[n][i]->getBuf(PARAMETER_GRADIENT);
-      if (grad) {
-        grad->zeroMem();
+    if (id == n || id == parameters_.size()) {
+      for (size_t i = 0; i < parameters_[n].size(); ++i) {
+        const VectorPtr& grad = parameters_[n][i]->getBuf(PARAMETER_GRADIENT);
+        if (grad) {
+          grad->zeroMem();
+        }
       }
     }
   }
 }
 
-void MKLDNNTester::clearBotDiffs() {
-  // dnn and ref
+void MKLDNNTester::clearBotDiffs(size_t id) {
+  CHECK_LE(id, dataLayers_.size());
   for (size_t n = 0; n < dataLayers_.size(); ++n) {
-    // all inputs layers
-    for (size_t i = 0; i < dataLayers_[n].size(); ++i) {
-      dataLayers_[n][i]->getOutputGrad()->zeroMem();
+    if (id == n || id == dataLayers_.size()) {
+      // clear inputs layers of this specific layer
+      for (size_t i = 0; i < dataLayers_[n].size(); ++i) {
+        dataLayers_[n][i]->getOutputGrad()->zeroMem();
+      }
     }
   }
 }
 
-void MKLDNNTester::clearBotDiffs(int n) {
-  CHECK_LT(n, NUM);
-  // all inputs layers
-  for (size_t i = 0; i < dataLayers_[n].size(); ++i) {
-    dataLayers_[n][i]->getOutputGrad()->zeroMem();
-  }
-}
-
-void MKLDNNTester::clearTopDatas() {
+void MKLDNNTester::clearTopDatas(size_t id) {
+  CHECK_LE(id, testLayers_.size());
   for (size_t i = 0; i < testLayers_.size(); ++i) {
-    testLayers_[i]->getOutputValue()->zeroMem();
+    if (id == i || id == testLayers_.size()) {
+      testLayers_[i]->getOutputValue()->zeroMem();
+    }
   }
 }
 
@@ -300,16 +304,24 @@ void MKLDNNTester::runOnce() {
   checkForward();
 
   // test backward
+  // simple updater
+  UpdateCallback updateCallback = [](Parameter* para) {
+    auto& grad = para->getBuf(PARAMETER_GRADIENT);
+    auto& value = para->getBuf(PARAMETER_VALUE);
+    real lr = 1e-3;
+    value->add(*grad, lr);
+  };
   randomTopDiffs();
-  dnnLayer_->backward(nullptr);
-  refLayer_->backward(nullptr);
+  dnnLayer_->backward(updateCallback);
+  refLayer_->backward(updateCallback);
   checkBackwardData();
   checkBackwardWgts();
 
   // clear buffers
   // ref code will addto the diff, dnn code will writeto it
-  // and clearTopDatas() and clearWgtDiffs() should be coverd by test layers
+  // and clearTopDatas(REF) should be coverd by ref layers
   clearBotDiffs(REF);
+  clearWgtDiffs(REF);
 }
 
 void MKLDNNTester::run(const TestConfig& dnn,
diff --git a/paddle/gserver/tests/MKLDNNTester.h b/paddle/gserver/tests/MKLDNNTester.h
index e55e4493ff..5ac885638c 100644
--- a/paddle/gserver/tests/MKLDNNTester.h
+++ b/paddle/gserver/tests/MKLDNNTester.h
@@ -18,6 +18,7 @@ limitations under the License. */
 #include <vector>
 #include "LayerGradUtil.h"
 #include "paddle/gserver/layers/MKLDNNBase.h"
+#include "paddle/gserver/layers/MKLDNNLayer.h"
 
 namespace paddle {
 
@@ -40,7 +41,8 @@ protected:
   vector<LayerMap> layerMaps_;
   vector<vector<ParameterPtr>> parameters_;
   vector<LayerPtr> testLayers_;
-  LayerPtr dnnLayer_, refLayer_;
+  LayerPtr refLayer_;
+  MKLDNNLayerPtr dnnLayer_;
 
   /// run some iterations, all the result should pass
   size_t iter_;
@@ -88,10 +90,10 @@ private:
   void checkBackwardData();
   void checkBackwardWgts();
 
-  void clearWgtDiffs();
-  void clearBotDiffs();
-  void clearBotDiffs(int n);  // clear specific layer
-  void clearTopDatas();
+  // clear specific layer, clear all when id equals NUM
+  void clearWgtDiffs(size_t id = NUM);
+  void clearBotDiffs(size_t id = NUM);
+  void clearTopDatas(size_t id = NUM);
 
   void printTopDatas();
   void printMatrix(const MatrixPtr& m);
diff --git a/paddle/gserver/tests/test_LayerGrad.cpp b/paddle/gserver/tests/test_LayerGrad.cpp
index 0e6be2df9e..090bde7b20 100644
--- a/paddle/gserver/tests/test_LayerGrad.cpp
+++ b/paddle/gserver/tests/test_LayerGrad.cpp
@@ -2302,26 +2302,27 @@ void test3DDeConvLayer(const string& type, bool trans, bool useGpu) {
   conv->set_stride(2);
   conv->set_stride_y(2);
   conv->set_stride_z(2);
-  conv->set_img_size(IMAGE_SIZE);
-  conv->set_img_size_y(IMAGE_SIZE_Y);
-  conv->set_img_size_z(IMAGE_SIZE_Z);
-  conv->set_output_x(imageSize(conv->img_size(),
+  conv->set_output_x(IMAGE_SIZE);
+  conv->set_output_y(IMAGE_SIZE_Y);
+  conv->set_output_z(IMAGE_SIZE_Z);
+
+  conv->set_img_size(imageSize(conv->output_x(),
                                conv->filter_size(),
                                conv->padding(),
                                conv->stride(),
                                true));
-  conv->set_output_y(imageSize(conv->img_size_y(),
-                               conv->filter_size_y(),
-                               conv->padding_y(),
-                               conv->stride_y(),
-                               true));
-  conv->set_output_z(imageSize(conv->img_size_z(),
-                               conv->filter_size_z(),
-                               conv->padding_z(),
-                               conv->stride_z(),
-                               true));
-  config.layerConfig.set_size(conv->output_x() * conv->output_y() *
-                              conv->output_z() * NUM_FILTERS);
+  conv->set_img_size_y(imageSize(conv->output_y(),
+                                 conv->filter_size_y(),
+                                 conv->padding_y(),
+                                 conv->stride_y(),
+                                 true));
+  conv->set_img_size_z(imageSize(conv->output_z(),
+                                 conv->filter_size_z(),
+                                 conv->padding_z(),
+                                 conv->stride_z(),
+                                 true));
+  config.layerConfig.set_size(conv->img_size() * conv->img_size_y() *
+                              conv->img_size_z() * NUM_FILTERS);
   conv->set_groups(1);
   conv->set_filter_channels(conv->channels() / conv->groups());
   config.inputDefs.push_back(
diff --git a/paddle/gserver/tests/test_MKLDNN.cpp b/paddle/gserver/tests/test_MKLDNN.cpp
index e1d2270df2..b593f65fe4 100644
--- a/paddle/gserver/tests/test_MKLDNN.cpp
+++ b/paddle/gserver/tests/test_MKLDNN.cpp
@@ -17,6 +17,7 @@ limitations under the License. */
 #include <vector>
 #include "MKLDNNTester.h"
 #include "ModelConfig.pb.h"
+#include "paddle/math/MathUtils.h"
 
 using namespace paddle;  // NOLINT
 
@@ -63,6 +64,145 @@ TEST(MKLDNNLayer, FcLayer) {
   testFcLayer({/*bs*/ 15, /*ic*/ 3, /*oc*/ 6, /*ih*/ 16, /*iw*/ 16});
 }
 
+struct testConvDesc {
+  int bs, gp;
+  int ic, ih, iw;
+  int oc, oh, ow;
+  int fh, fw;
+  int ph, pw;
+  int sh, sw;
+  int dh, dw;
+};
+
+void testConvLayer(const testConvDesc& pm) {
+  const std::string compareTypes[] = {"mkldnn_conv", "exconv"};
+  TestConfig cfg;
+  cfg.layerConfig.set_type(compareTypes[0]);
+  cfg.layerConfig.set_num_filters(pm.oc);
+  cfg.layerConfig.set_size(pm.oc * pm.oh * pm.ow);
+  // cfg.layerConfig.set_partial_sum(1); // TODO: check it
+  cfg.layerConfig.set_shared_biases(true);
+  cfg.inputDefs.push_back(
+      {INPUT_DATA,
+       "layer_0",
+       /* size of input layer= */ size_t(pm.ic * pm.ih * pm.iw),
+       /* size of weight= */ size_t(pm.oc * pm.ic * pm.fh * pm.fw / pm.gp)});
+  LayerInputConfig* input = cfg.layerConfig.add_inputs();
+  ConvConfig* conv = input->mutable_conv_conf();
+  conv->set_groups(pm.gp);
+  conv->set_img_size(pm.iw);
+  conv->set_img_size_y(pm.ih);
+  conv->set_output_x(pm.ow);
+  conv->set_output_y(pm.oh);
+  conv->set_filter_size(pm.fw);
+  conv->set_filter_size_y(pm.fh);
+  conv->set_channels(pm.ic);
+  conv->set_padding(pm.pw);
+  conv->set_padding_y(pm.ph);
+  conv->set_stride(pm.sw);
+  conv->set_stride_y(pm.sh);
+  conv->set_dilation(pm.dw);
+  conv->set_dilation_y(pm.dh);
+  conv->set_caffe_mode(true);
+  conv->set_filter_channels(conv->channels() / conv->groups());
+  CHECK_EQ(conv->filter_channels() * pm.gp, conv->channels())
+      << "it is indivisible";
+
+  int fh = (pm.fh - 1) * pm.dh + 1;
+  int fw = (pm.fw - 1) * pm.dw + 1;
+  int ow = outputSize(pm.iw, fw, pm.pw, pm.sw, true);
+  int oh = outputSize(pm.ih, fh, pm.ph, pm.sh, true);
+  CHECK_EQ(ow, pm.ow) << "output size check failed";
+  CHECK_EQ(oh, pm.oh) << "output size check failed";
+
+  MKLDNNTester tester;
+  for (auto biasSize : {pm.oc, 0}) {
+    cfg.biasSize = biasSize;
+    TestConfig ref = cfg;
+    ref.layerConfig.set_type(compareTypes[1]);
+    for (auto bs : {pm.bs, 1}) {
+      tester.run(cfg, ref, bs, pm.ih, pm.iw);
+    }
+  }
+}
+
+TEST(MKLDNNLayer, ConvLayer) {
+  /* bs, gp, ic, ih, iw, oc, oh, ow, fh, fw, ph, pw, sh, sw, dh, dw */
+  testConvLayer({2, 1, 3, 32, 32, 16, 32, 32, 3, 3, 1, 1, 1, 1, 1, 1});
+  testConvLayer({2, 1, 8, 16, 16, 8, 16, 16, 3, 3, 1, 1, 1, 1, 1, 1});
+  testConvLayer({3, 1, 16, 32, 32, 3, 32, 32, 3, 3, 1, 1, 1, 1, 1, 1});
+  testConvLayer({8, 1, 16, 18, 18, 32, 18, 18, 3, 3, 1, 1, 1, 1, 1, 1});
+  testConvLayer({16, 1, 1, 42, 31, 32, 23, 11, 4, 5, 3, 2, 2, 3, 1, 1});
+  testConvLayer({2, 1, 8, 16, 16, 8, 8, 8, 3, 3, 1, 1, 2, 2, 1, 1});
+  testConvLayer({3, 1, 8, 13, 13, 8, 7, 7, 3, 3, 1, 1, 2, 2, 1, 1});
+  // with groups
+  testConvLayer({2, 2, 4, 5, 5, 8, 5, 5, 3, 3, 1, 1, 1, 1, 1, 1});
+  testConvLayer({2, 3, 3, 5, 5, 3, 5, 5, 3, 3, 1, 1, 1, 1, 1, 1});
+  testConvLayer({4, 4, 16, 3, 3, 16, 3, 3, 3, 3, 1, 1, 1, 1, 1, 1});
+}
+
+struct testPoolDesc {
+  int bs, ch;  // input channel and output channel are the same
+  int ih, iw;
+  int oh, ow;
+  int fh, fw;
+  int ph, pw;
+  int sh, sw;
+};
+
+void testPoolLayer(const testPoolDesc& pm) {
+  const std::string compareTypes[] = {"mkldnn_pool", "pool"};
+  TestConfig cfg;
+  cfg.layerConfig.set_type(compareTypes[0]);
+  cfg.layerConfig.set_size(pm.ch * pm.oh * pm.ow);
+  cfg.inputDefs.push_back(
+      {INPUT_DATA,
+       "layer_0",
+       /* size of input layer= */ size_t(pm.ch * pm.ih * pm.iw),
+       0});
+  LayerInputConfig* input = cfg.layerConfig.add_inputs();
+  PoolConfig* pool = input->mutable_pool_conf();
+  // pool->set_pool_type(poolType);
+  pool->set_channels(pm.ch);
+  pool->set_img_size(pm.iw);
+  pool->set_img_size_y(pm.ih);
+  pool->set_output_x(pm.ow);
+  pool->set_output_y(pm.oh);
+  pool->set_size_x(pm.fw);
+  pool->set_size_y(pm.fh);
+  pool->set_padding(pm.pw);
+  pool->set_padding_y(pm.ph);
+  pool->set_stride(pm.sw);
+  pool->set_stride_y(pm.sh);
+
+  int oh = outputSize(pm.ih, pm.fh, pm.ph, pm.sh, false);
+  int ow = outputSize(pm.iw, pm.fw, pm.pw, pm.sw, false);
+  CHECK_EQ(ow, pm.ow) << "output size check failed";
+  CHECK_EQ(oh, pm.oh) << "output size check failed";
+
+  MKLDNNTester tester;
+  for (auto type : {"max-projection", "avg-projection"}) {
+    pool->set_pool_type(type);
+    TestConfig ref = cfg;
+    ref.layerConfig.set_type(compareTypes[1]);
+    for (auto bs : {pm.bs, 1}) {
+      tester.run(cfg, ref, bs, pm.ih, pm.iw);
+    }
+  }
+}
+
+TEST(MkldnnLayer, PoolLayer) {
+  /* bs, ch, ih, iw, oh, ow, fh, fw, ph, pw, sh, sw*/
+  testPoolLayer({2, 1, 4, 4, 2, 2, 3, 3, 0, 0, 2, 2});
+  testPoolLayer({10, 8, 16, 16, 8, 8, 2, 2, 0, 0, 2, 2});
+  testPoolLayer({4, 2, 5, 5, 3, 3, 3, 3, 1, 1, 2, 2});
+  testPoolLayer({8, 16, 56, 56, 28, 28, 3, 3, 0, 0, 2, 2});
+  testPoolLayer({8, 16, 14, 14, 7, 7, 3, 3, 0, 0, 2, 2});
+  testPoolLayer({4, 16, 7, 7, 1, 1, 7, 7, 0, 0, 1, 1});
+  testPoolLayer({4, 2, 5, 5, 3, 3, 5, 5, 1, 1, 1, 1});
+  testPoolLayer({2, 8, 56, 56, 29, 29, 3, 3, 1, 1, 2, 2});
+}
+
 // TODO(TJ): add branch test
 
 int main(int argc, char** argv) {
diff --git a/paddle/math/BaseMatrix.cu b/paddle/math/BaseMatrix.cu
index 5435808fb7..53dd538360 100644
--- a/paddle/math/BaseMatrix.cu
+++ b/paddle/math/BaseMatrix.cu
@@ -17,6 +17,7 @@ limitations under the License. */
 #include <cmath>
 #include "BaseMatrix.h"
 #include "MathFunctions.h"
+#include "NEONFunctions.h"
 #include "SIMDFunctions.h"
 #include "hl_matrix_apply.cuh"
 #include "hl_matrix_base.cuh"
@@ -666,6 +667,13 @@ void BaseMatrixT<T>::relu(BaseMatrixT& b) {
   applyBinary(binary::Relu<T>(), b);
 }
 
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+template <>
+void BaseMatrixT<float>::relu(BaseMatrixT& b) {
+  neon::relu(data_, b.data_, height_ * width_);
+}
+#endif
+
 DEFINE_MATRIX_BINARY_OP(ReluDerivative, a *= (b > 0.0f ? 1.0f : 0.0f));
 template <class T>
 void BaseMatrixT<T>::reluDerivative(BaseMatrixT& b) {
diff --git a/paddle/math/MKLDNNMatrix.cpp b/paddle/math/MKLDNNMatrix.cpp
index 0a355e2644..0778bb63b7 100644
--- a/paddle/math/MKLDNNMatrix.cpp
+++ b/paddle/math/MKLDNNMatrix.cpp
@@ -33,14 +33,12 @@ MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m, memory::primitive_desc pd) {
     size_t width = cnts / dims[0];
     m = Matrix::create(height, width, false, false);
   }
-
   CHECK(m) << " Matrix should not be empty";
+
   CpuMatrixPtr cpuMatrix = std::dynamic_pointer_cast<CpuMatrix>(m);
   CHECK(cpuMatrix) << "Only support create from CPU matrix yet";
-
-  CHECK_EQ(cnts, m->getElementCnt()) << "Count size does not match";
-  return std::make_shared<MKLDNNMatrix>(
-      m->getData(), m->getHeight(), m->getWidth(), pd);
+  CHECK_EQ(cpuMatrix->getElementCnt(), cnts) << "Count size does not match";
+  return std::make_shared<MKLDNNMatrix>(cpuMatrix, pd);
 }
 
 MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m,
@@ -51,6 +49,27 @@ MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m,
   return create(m, memory::primitive_desc(memory::desc(dims, dtype, fmt), eg));
 }
 
+std::shared_ptr<reorder> MKLDNNMatrix::createReorder(const MKLDNNMatrixPtr& src,
+                                                     const MKLDNNMatrixPtr& dst,
+                                                     bool checkData) {
+  if (src == dst || src->getPrimitiveDesc() == dst->getPrimitiveDesc()) {
+    return nullptr;
+  }
+
+  if (checkData && (src->getData() == dst->getData())) {
+    LOG(FATAL) << "can not create reorder with inplace data";
+    return nullptr;
+  }
+
+  memory::dims srcDims = src->getDims();
+  memory::dims dstDims = dst->getDims();
+  CHECK_EQ(srcDims.size(), dstDims.size());
+  for (size_t i = 0; i < srcDims.size(); ++i) {
+    CHECK_EQ(srcDims[i], dstDims[i]);
+  }
+  return std::make_shared<reorder>(*src, *dst);
+}
+
 void MKLDNNMatrix::reorderDataFrom(const MKLDNNMatrixPtr& m,
                                    memory::format srcFmt,
                                    memory::dims targetDim) {
@@ -138,7 +157,7 @@ void MKLDNNMatrix::downSpatial() {
       mkldnn_primitive_create(&result, pd.get(), nullptr, nullptr),
       "could not create a memory primitive");
   reset(result);
-  set_data_handle(getData());
+  set_data_handle(data_);
 }
 
 }  // namespace paddle
diff --git a/paddle/math/MKLDNNMatrix.h b/paddle/math/MKLDNNMatrix.h
index e50f698b49..c843115eb9 100644
--- a/paddle/math/MKLDNNMatrix.h
+++ b/paddle/math/MKLDNNMatrix.h
@@ -30,11 +30,10 @@ typedef std::shared_ptr<MKLDNNMatrix> MKLDNNMatrixPtr;
  */
 class MKLDNNMatrix : public CpuMatrix, public mkldnn::memory {
 public:
-  MKLDNNMatrix(real* data,
-               size_t height,
-               size_t width,
-               mkldnn::memory::primitive_desc pd)
-      : CpuMatrix(data, height, width, false), mkldnn::memory(pd, data) {}
+  MKLDNNMatrix(CpuMatrixPtr m, mkldnn::memory::primitive_desc pd)
+      : CpuMatrix(m->getData(), m->getHeight(), m->getWidth(), false),
+        mkldnn::memory(pd, m->getData()),
+        m_(m) {}
 
   ~MKLDNNMatrix() {}
 
@@ -53,6 +52,32 @@ public:
       mkldnn::engine& eg,
       mkldnn::memory::data_type dtype = mkldnn::memory::data_type::f32);
 
+  /**
+   * Create Memory descriptor.
+   * default with any format and f32 dtype
+   */
+  static mkldnn::memory::desc createMemoryDesc(
+      const mkldnn::memory::dims& dims,
+      const mkldnn::memory::format& fmt = mkldnn::memory::format::any,
+      const mkldnn::memory::data_type& dtype = mkldnn::memory::data_type::f32) {
+    return mkldnn::memory::desc(dims, dtype, fmt);
+  }
+
+  /**
+   * Create reorder primitive.
+   * Create a mkldnn::reorder handle for converting src MKLDNNMatrix to dst.
+   * checkData: whether to check the data handle of src and dst.
+   *            if true, it will check the data and do not allow them equal;
+   *            otherwise, it will not check them, then the reorder created
+   *            may have inplace buffer.
+   *            Do not set false, if you can not guarantee the inplace logical
+   *            would work with your reorder.
+   */
+  static std::shared_ptr<mkldnn::reorder> createReorder(
+      const MKLDNNMatrixPtr& src,
+      const MKLDNNMatrixPtr& dst,
+      bool checkData = true);
+
 public:
   /**
    * Reorder this MKLDNNMatrix from other format.
@@ -81,11 +106,29 @@ public:
   void downSpatial();
 
   /**
-   * Update the memory data handle.
+   * set the memory data handle.
    * Caution: This will not check the buffer size of the data,
    *          it should be coverd by user.
    */
-  void updateData(void* data) { set_data_handle(data); }
+  void setData(real* data) {
+    set_data_handle(data);
+    CpuMatrix::setData(data);
+    m_.reset();
+  }
+
+  /**
+   * override Matrix::getData
+   * check data before return
+   */
+  real* getData() override {
+    CHECK_EQ((void*)data_, get_data_handle());
+    return data_;
+  }
+
+  const real* getData() const override {
+    CHECK_EQ((void*)data_, get_data_handle());
+    return data_;
+  }
 
   /**
    * Get primitive descriptor.
@@ -143,6 +186,10 @@ protected:
                    memory::format srcFmt,
                    memory::format dstFmt,
                    memory::dims dm);
+
+private:
+  // save the CpuMatrixPtr in case the buffer released outside
+  CpuMatrixPtr m_;
 };
 
 }  // namespace paddle
diff --git a/paddle/math/Matrix.cpp b/paddle/math/Matrix.cpp
index 4a2132c8d1..0023b4d0f5 100644
--- a/paddle/math/Matrix.cpp
+++ b/paddle/math/Matrix.cpp
@@ -1033,17 +1033,15 @@ void GpuMatrix::maxPoolForward(Matrix& inputMat,
 
   real* inputData = inputMat.getData();
   size_t frameNum = inputMat.getHeight();
-  size_t width = imgSizeW;
-  size_t height = imgSizeH;
-  CHECK(height * width * channels == inputMat.getWidth());
+  CHECK(imgSizeH * imgSizeW * channels == inputMat.getWidth());
   CHECK(height_ == inputMat.getHeight());
   CHECK(width_ == outputH * outputW * channels);
 
   hl_maxpool_forward(frameNum,
                      inputData,
                      channels,
-                     height,
-                     width,
+                     imgSizeH,
+                     imgSizeW,
                      outputH,
                      outputW,
                      sizeX,
@@ -1080,11 +1078,8 @@ void GpuMatrix::maxPoolBackward(Matrix& inputMat,
   real* outDiff = outGrad.getData();
   size_t frameNum = inputMat.getHeight();
   size_t channels = outV.getWidth() / outputH / outputW;
-  size_t width = imgSizeW;
-  size_t height = imgSizeH;
-  CHECK(height * width * channels == inputMat.getWidth());
+  CHECK(imgSizeH * imgSizeW * channels == inputMat.getWidth());
   CHECK(height_ == inputMat.getHeight());
-  CHECK(width_ == width * height * channels);
   CHECK(outGrad.getHeight() == outV.getHeight() &&
         outGrad.getWidth() == outV.getWidth());
 
@@ -1093,8 +1088,8 @@ void GpuMatrix::maxPoolBackward(Matrix& inputMat,
                       outData,
                       outDiff,
                       channels,
-                      height,
-                      width,
+                      imgSizeH,
+                      imgSizeW,
                       outputH,
                       outputW,
                       sizeX,
@@ -1125,17 +1120,15 @@ void GpuMatrix::avgPoolForward(Matrix& inputMat,
 
   real* inputData = inputMat.getData();
   size_t frameNum = inputMat.getHeight();
-  size_t height = imgSizeH;
-  size_t width = imgSizeW;
-  CHECK(height * width * channels == inputMat.getWidth());
+  CHECK(imgSizeH * imgSizeW * channels == inputMat.getWidth());
   CHECK(height_ == inputMat.getHeight());
   CHECK(width_ == outputH * outputW * channels);
 
   hl_avgpool_forward(frameNum,
                      inputData,
                      channels,
-                     height,
-                     width,
+                     imgSizeH,
+                     imgSizeW,
                      outputH,
                      outputW,
                      sizeX,
@@ -1166,17 +1159,15 @@ void GpuMatrix::avgPoolBackward(Matrix& outGrad,
   real* outDiff = outGrad.getData();
   size_t frameNum = outGrad.getHeight();
   size_t channels = outGrad.getWidth() / outputH / outputW;
-  size_t height = imgSizeH;
-  size_t width = imgSizeW;
-  CHECK(height * width * channels == width_);
+  CHECK(imgSizeH * imgSizeW * channels == width_);
   CHECK(height_ == outGrad.getHeight());
   CHECK(outGrad.getWidth() == outputH * outputW * channels);
 
   hl_avgpool_backward(frameNum,
                       outDiff,
                       channels,
-                      height,
-                      width,
+                      imgSizeH,
+                      imgSizeW,
                       outputH,
                       outputW,
                       sizeX,
@@ -1214,19 +1205,16 @@ void GpuMatrix::maxPool3DForward(Matrix& inputMat,
   real* inputData = inputMat.getData();
   real* maxPoolIdxData = maxPoolIdx.getData();
   size_t num = inputMat.getHeight();
-  size_t width = imgSizeW;
-  size_t height = imgSizeH;
-  size_t depth = imgSizeD;
-  CHECK(depth * height * width * channels == inputMat.getWidth());
+  CHECK(imgSizeD * imgSizeH * imgSizeW * channels == inputMat.getWidth());
   CHECK(height_ == inputMat.getHeight());
   CHECK(width_ == outputD * outputH * outputW * channels);
 
   hl_maxpool3D_forward(num,
                        inputData,
                        channels,
-                       depth,
-                       height,
-                       width,
+                       imgSizeD,
+                       imgSizeH,
+                       imgSizeW,
                        outputD,
                        outputH,
                        outputW,
@@ -1269,20 +1257,16 @@ void GpuMatrix::maxPool3DBackward(Matrix& outGrad,
   real* maxPoolIdxData = maxPoolIdx.getData();
   size_t frameNum = getHeight();
   size_t channels = outGrad.getWidth() / outputD / outputH / outputW;
-  size_t width = imgSizeW;
-  size_t height = imgSizeH;
-  size_t depth = imgSizeD;
-  CHECK(depth * height * width * channels == getWidth());
-  CHECK(width_ == depth * width * height * channels);
+  CHECK(imgSizeD * imgSizeH * imgSizeW * channels == getWidth());
   CHECK(outGrad.getHeight() == maxPoolIdx.getHeight() &&
         outGrad.getWidth() == maxPoolIdx.getWidth());
 
   hl_maxpool3D_backward(frameNum,
                         outDiff,
                         channels,
-                        depth,
-                        height,
-                        width,
+                        imgSizeD,
+                        imgSizeH,
+                        imgSizeW,
                         outputD,
                         outputH,
                         outputW,
@@ -1323,19 +1307,16 @@ void GpuMatrix::avgPool3DForward(Matrix& inputMat,
 
   real* inputData = inputMat.getData();
   size_t frameNum = inputMat.getHeight();
-  size_t height = imgSizeH;
-  size_t width = imgSizeW;
-  size_t depth = imgSizeD;
-  CHECK(depth * height * width * channels == inputMat.getWidth());
+  CHECK(imgSizeD * imgSizeH * imgSizeW * channels == inputMat.getWidth());
   CHECK(height_ == inputMat.getHeight());
   CHECK(width_ == outputD * outputH * outputW * channels);
 
   hl_avgpool3D_forward(frameNum,
                        inputData,
                        channels,
-                       depth,
-                       height,
-                       width,
+                       imgSizeD,
+                       imgSizeH,
+                       imgSizeW,
                        outputD,
                        outputH,
                        outputW,
@@ -1375,19 +1356,16 @@ void GpuMatrix::avgPool3DBackward(Matrix& outGrad,
   real* outDiff = outGrad.getData();
   size_t frameNum = outGrad.getHeight();
   size_t channels = outGrad.getWidth() / outputD / outputH / outputW;
-  size_t height = imgSizeH;
-  size_t width = imgSizeW;
-  size_t depth = imgSizeD;
-  CHECK(depth * height * width * channels == width_);
+  CHECK(imgSizeD * imgSizeH * imgSizeW * channels == width_);
   CHECK(height_ == outGrad.getHeight());
   CHECK(outGrad.getWidth() == outputD * outputH * outputW * channels);
 
   hl_avgpool3D_backward(frameNum,
                         outDiff,
                         channels,
-                        depth,
-                        height,
-                        width,
+                        imgSizeD,
+                        imgSizeH,
+                        imgSizeW,
                         outputD,
                         outputH,
                         outputW,
@@ -1999,11 +1977,11 @@ void CpuMatrix::maxPoolForward(Matrix& inputMat,
   real* inputData = inputMat.getData();
   real* outData = data_;
   size_t num = inputMat.getHeight();
-  size_t inWidth = imgSizeW;
-  size_t inHeight = imgSizeH;
-  CHECK(inHeight * inWidth == inputMat.getWidth() / channels);
+  size_t inLength = imgSizeH * imgSizeW;
+  size_t outLength = outputH * outputW;
+  CHECK(inLength == inputMat.getWidth() / channels);
   CHECK_EQ(num, this->getHeight());
-  CHECK_EQ(channels * outputH * outputW, this->getWidth());
+  CHECK_EQ(channels * outLength, this->getWidth());
   size_t outStride = getStride();
 
   /* initialize the data_ */
@@ -2020,24 +1998,24 @@ void CpuMatrix::maxPoolForward(Matrix& inputMat,
     }
     for (size_t c = 0; c < channels; ++c) {  // channel by channel
       for (size_t ph = 0; ph < outputH; ++ph) {
+        int hstart = ph * strideH - paddingH;
+        int hend = std::min(hstart + sizeY, imgSizeH);
+        hstart = std::max(hstart, 0);
         for (size_t pw = 0; pw < outputW; ++pw) {
-          int hstart = ph * strideH - paddingH;
           int wstart = pw * strideW - paddingW;
-          int hend = std::min(hstart + sizeY, inHeight);
-          int wend = std::min(wstart + sizeX, inWidth);
-          hstart = std::max(hstart, 0);
+          int wend = std::min(wstart + sizeX, imgSizeW);
           wstart = std::max(wstart, 0);
           for (int h = hstart; h < hend; ++h) {
             for (int w = wstart; w < wend; ++w) {
-              outData[ph * outputW + pw] = std::max(outData[ph * outputW + pw],
-                                                    inputData[h * inWidth + w]);
+              outData[ph * outputW + pw] = std::max(
+                  outData[ph * outputW + pw], inputData[h * imgSizeW + w]);
             }
           }
         }
       }
       // compute offset
-      inputData += inHeight * inWidth;
-      outData += outputH * outputW;
+      inputData += inLength;
+      outData += outLength;
     }
   }
 }
@@ -2058,8 +2036,10 @@ void CpuMatrix::maxPoolBackward(Matrix& image,
                                 size_t paddingH,
                                 size_t paddingW) {
   size_t num = image.getHeight();
-  size_t channels = size_t(width_ / imgSizeH / imgSizeW);
-  CHECK(image.getWidth() == imgSizeH * imgSizeW * channels);
+  size_t inLength = imgSizeH * imgSizeW;
+  size_t outLength = outputH * outputW;
+  size_t channels = size_t(width_ / inLength);
+  CHECK(image.getWidth() == inLength * channels);
   CHECK(image.getHeight() == height_ && image.getWidth() == width_);
   CHECK(outV.getHeight() == outGrad.getHeight() &&
         outV.getWidth() == outGrad.getWidth());
@@ -2080,12 +2060,12 @@ void CpuMatrix::maxPoolBackward(Matrix& image,
     }
     for (size_t c = 0; c < channels; ++c) {
       for (size_t ph = 0; ph < outputH; ++ph) {
+        int hstart = ph * strideH - paddingH;
+        int hend = std::min(hstart + sizeY, imgSizeH);
+        hstart = std::max(hstart, 0);
         for (size_t pw = 0; pw < outputW; ++pw) {
-          int hstart = ph * strideH - paddingH;
           int wstart = pw * strideW - paddingW;
-          int hend = std::min(hstart + sizeY, imgSizeH);
           int wend = std::min(wstart + sizeX, imgSizeW);
-          hstart = std::max(hstart, 0);
           wstart = std::max(wstart, 0);
           for (int h = hstart; h < hend; ++h) {
             for (int w = wstart; w < wend; ++w) {
@@ -2098,10 +2078,10 @@ void CpuMatrix::maxPoolBackward(Matrix& image,
         }
       }
       // offset
-      inData += imgSizeH * imgSizeW;
-      tgtGrad += imgSizeH * imgSizeW;
-      otData += outputH * outputW;
-      otGrad += outputH * outputW;
+      inData += inLength;
+      tgtGrad += inLength;
+      otData += outLength;
+      otGrad += outLength;
     }
   }
 }
@@ -2120,10 +2100,10 @@ void CpuMatrix::avgPoolForward(Matrix& input,
                                size_t paddingW) {
   // The main loop
   size_t num = input.getHeight();
-  size_t inHeight = imgSizeH;
-  size_t inWidth = imgSizeW;
-  CHECK(inHeight * inWidth * channels == input.getWidth());
-  CHECK(outputH * outputW * channels * num == height_ * width_);
+  size_t inLength = imgSizeH * imgSizeW;
+  size_t outLength = outputH * outputW;
+  CHECK(inLength * channels == input.getWidth());
+  CHECK(outLength * channels * num == height_ * width_);
   real* tgtData = data_;
   real* inData = input.getData();
 
@@ -2133,30 +2113,27 @@ void CpuMatrix::avgPoolForward(Matrix& input,
     }
     for (size_t c = 0; c < channels; ++c) {
       for (size_t ph = 0; ph < outputH; ++ph) {
+        int hstart = ph * strideH - paddingH;
+        int hend = std::min(hstart + sizeY, imgSizeH);
+        hstart = std::max(hstart, 0);
         for (size_t pw = 0; pw < outputW; ++pw) {
-          int hstart = ph * strideH - paddingH;
           int wstart = pw * strideW - paddingW;
-          int hend = std::min(hstart + sizeY, inHeight + paddingH);
-          int wend = std::min(wstart + sizeX, inWidth + paddingW);
-          int poolSize = (hend - hstart) * (wend - wstart);
-          hstart = std::max(hstart, 0);
+          int wend = std::min(wstart + sizeX, imgSizeW);
           wstart = std::max(wstart, 0);
-          hend = std::min(hend, static_cast<int>(inHeight));
-          wend = std::min(wend, static_cast<int>(inWidth));
-
-          CHECK(poolSize);
           tgtData[ph * outputW + pw] = 0;  // clear
           for (int h = hstart; h < hend; ++h) {
             for (int w = wstart; w < wend; ++w) {
-              tgtData[ph * outputW + pw] += inData[h * inWidth + w];
+              tgtData[ph * outputW + pw] += inData[h * imgSizeW + w];
             }
           }
+          int poolSize = (hend - hstart) * (wend - wstart);
+          CHECK(poolSize);
           tgtData[ph * outputW + pw] /= poolSize;
         }
       }
       // compute offset
-      inData += inHeight * inWidth;
-      tgtData += outputH * outputW;
+      inData += inLength;
+      tgtData += outLength;
     }
   }
 }
@@ -2176,7 +2153,9 @@ void CpuMatrix::avgPoolBackward(Matrix& input,
                                 size_t paddingW) {
   size_t num = input.getHeight();
   size_t channels = input.getWidth() / outputH / outputW;
-  CHECK(imgSizeH * imgSizeW * channels == getWidth());
+  size_t inLength = imgSizeH * imgSizeW;
+  size_t outLength = outputH * outputW;
+  CHECK(inLength * channels == getWidth());
   real* inData = input.getData();
   real* outData = getData();
 
@@ -2186,16 +2165,14 @@ void CpuMatrix::avgPoolBackward(Matrix& input,
     }
     for (size_t c = 0; c < channels; ++c) {
       for (size_t ph = 0; ph < outputH; ++ph) {
+        int hstart = ph * strideH - paddingH;
+        int hend = std::min(hstart + sizeY, imgSizeH);
+        hstart = std::max(hstart, 0);
         for (size_t pw = 0; pw < outputW; ++pw) {
-          int hstart = ph * strideH - paddingH;
           int wstart = pw * strideW - paddingW;
-          int hend = std::min(hstart + sizeY, imgSizeH + paddingH);
-          int wend = std::min(wstart + sizeX, imgSizeW + paddingW);
-          int poolSize = (hend - hstart) * (wend - wstart);
-          hstart = std::max(hstart, 0);
+          int wend = std::min(wstart + sizeX, imgSizeW);
           wstart = std::max(wstart, 0);
-          hend = std::min(hend, static_cast<int>(imgSizeH));
-          wend = std::min(wend, static_cast<int>(imgSizeW));
+          int poolSize = (hend - hstart) * (wend - wstart);
           CHECK(poolSize);
 
           for (int h = hstart; h < hend; ++h) {
@@ -2206,8 +2183,8 @@ void CpuMatrix::avgPoolBackward(Matrix& input,
         }
       }
       // offset
-      outData += imgSizeH * imgSizeW;
-      inData += outputH * outputW;
+      outData += inLength;
+      inData += outLength;
     }
   }
 }
@@ -2234,12 +2211,11 @@ void CpuMatrix::maxPool3DForward(Matrix& inputMat,
   real* outData = getData();
   real* maxPoolIdxData = maxPoolIdx.getData();
   size_t num = inputMat.getHeight();
-  size_t inWidth = imgSizeW;
-  size_t inHeight = imgSizeH;
-  size_t inDepth = imgSizeD;
-  CHECK(inHeight * inWidth * inDepth == inputMat.getWidth() / channels);
+  size_t inLength = imgSizeH * imgSizeW * imgSizeD;
+  size_t outLength = outputH * outputW * outputD;
+  CHECK(inLength == inputMat.getWidth() / channels);
   CHECK_EQ(num, this->getHeight());
-  CHECK_EQ(channels * outputH * outputW * outputD, this->getWidth());
+  CHECK_EQ(channels * outLength, this->getWidth());
   size_t outStride = getStride();
 
   /* initialize the data_ */
@@ -2258,16 +2234,16 @@ void CpuMatrix::maxPool3DForward(Matrix& inputMat,
     }
     for (size_t c = 0; c < channels; ++c) {  // channel by channel
       for (size_t pd = 0; pd < outputD; ++pd) {
+        int dstart = pd * strideD - paddingD;
+        int dend = std::min(dstart + sizeZ, imgSizeD);
+        dstart = std::max(dstart, 0);
         for (size_t ph = 0; ph < outputH; ++ph) {
+          int hstart = ph * strideH - paddingH;
+          int hend = std::min(hstart + sizeY, imgSizeH);
+          hstart = std::max(hstart, 0);
           for (size_t pw = 0; pw < outputW; ++pw) {
-            int dstart = pd * strideD - paddingD;
-            int hstart = ph * strideH - paddingH;
             int wstart = pw * strideW - paddingW;
-            int dend = std::min(dstart + sizeZ, inDepth);
-            int hend = std::min(hstart + sizeY, inHeight);
-            int wend = std::min(wstart + sizeX, inWidth);
-            dstart = std::max(dstart, 0);
-            hstart = std::max(hstart, 0);
+            int wend = std::min(wstart + sizeX, imgSizeW);
             wstart = std::max(wstart, 0);
             int maxIdx = -1;
             real maxOutData = outData[(pd * outputH + ph) * outputW + pw];
@@ -2275,9 +2251,9 @@ void CpuMatrix::maxPool3DForward(Matrix& inputMat,
               for (int h = hstart; h < hend; ++h) {
                 for (int w = wstart; w < wend; ++w) {
                   if (maxOutData <
-                      inputData[(d * inHeight + h) * inWidth + w]) {
-                    maxOutData = inputData[(d * inHeight + h) * inWidth + w];
-                    maxIdx = (d * inHeight + h) * inWidth + w;
+                      inputData[(d * imgSizeH + h) * imgSizeW + w]) {
+                    maxOutData = inputData[(d * imgSizeH + h) * imgSizeW + w];
+                    maxIdx = (d * imgSizeH + h) * imgSizeW + w;
                   }
                 }
               }
@@ -2288,9 +2264,9 @@ void CpuMatrix::maxPool3DForward(Matrix& inputMat,
         }
       }
       // compute offset
-      inputData += inDepth * inHeight * inWidth;
-      outData += outputD * outputH * outputW;
-      maxPoolIdxData += outputD * outputH * outputW;
+      inputData += inLength;
+      outData += outLength;
+      maxPoolIdxData += outLength;
     }
   }
 }
@@ -2315,7 +2291,9 @@ void CpuMatrix::maxPool3DBackward(Matrix& outGrad,
                                   real scaleTargets,
                                   real scaleOutput) {
   size_t num = getHeight();
-  size_t channels = size_t(width_ / imgSizeD / imgSizeH / imgSizeW);
+  size_t inLength = imgSizeH * imgSizeW * imgSizeD;
+  size_t outLength = outputH * outputW * outputD;
+  size_t channels = size_t(width_ / inLength);
   CHECK(maxPoolIdx.getHeight() == outGrad.getHeight() &&
         maxPoolIdx.getWidth() == outGrad.getWidth());
 
@@ -2341,9 +2319,9 @@ void CpuMatrix::maxPool3DBackward(Matrix& outGrad,
         }
       }
       // offset
-      tgtGrad += imgSizeD * imgSizeH * imgSizeW;
-      otGrad += outputD * outputH * outputW;
-      maxPoolIdxData += outputD * outputH * outputW;
+      tgtGrad += inLength;
+      otGrad += outLength;
+      maxPoolIdxData += outLength;
     }
   }
 }
@@ -2367,11 +2345,10 @@ void CpuMatrix::avgPool3DForward(Matrix& input,
                                  size_t paddingW) {
   // The main loop
   size_t num = input.getHeight();
-  size_t inDepth = imgSizeD;
-  size_t inHeight = imgSizeH;
-  size_t inWidth = imgSizeW;
-  CHECK(inDepth * inHeight * inWidth * channels == input.getWidth());
-  CHECK(outputD * outputH * outputW * channels * num == height_ * width_);
+  size_t inLength = imgSizeH * imgSizeW * imgSizeD;
+  size_t outLength = outputH * outputW * outputD;
+  CHECK(inLength * channels == input.getWidth());
+  CHECK(outLength * channels * num == height_ * width_);
   real* tgtData = getData();
   real* inData = input.getData();
 
@@ -2381,39 +2358,36 @@ void CpuMatrix::avgPool3DForward(Matrix& input,
     }
     for (size_t c = 0; c < channels; ++c) {
       for (size_t pd = 0; pd < outputD; ++pd) {
+        int dstart = pd * strideD - paddingD;
+        int dend = std::min(dstart + sizeZ, imgSizeD);
+        dstart = std::max(dstart, 0);
         for (size_t ph = 0; ph < outputH; ++ph) {
+          int hstart = ph * strideH - paddingH;
+          int hend = std::min(hstart + sizeY, imgSizeH);
+          hstart = std::max(hstart, 0);
           for (size_t pw = 0; pw < outputW; ++pw) {
-            int dstart = pd * strideD - paddingD;
-            int hstart = ph * strideH - paddingH;
             int wstart = pw * strideW - paddingW;
-            int dend = std::min(dstart + sizeZ, inDepth + paddingD);
-            int hend = std::min(hstart + sizeY, inHeight + paddingH);
-            int wend = std::min(wstart + sizeX, inWidth + paddingW);
-            int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart);
-            dstart = std::max(dstart, 0);
-            hstart = std::max(hstart, 0);
+            int wend = std::min(wstart + sizeX, imgSizeW);
             wstart = std::max(wstart, 0);
-            dend = std::min(dend, static_cast<int>(inDepth));
-            hend = std::min(hend, static_cast<int>(inHeight));
-            wend = std::min(wend, static_cast<int>(inWidth));
 
-            CHECK(poolSize);
             tgtData[(pd * outputH + ph) * outputW + pw] = 0;  // clear
             for (int d = dstart; d < dend; ++d) {
               for (int h = hstart; h < hend; ++h) {
                 for (int w = wstart; w < wend; ++w) {
                   tgtData[(pd * outputH + ph) * outputW + pw] +=
-                      inData[(d * inHeight + h) * inWidth + w];
+                      inData[(d * imgSizeH + h) * imgSizeW + w];
                 }
               }
             }
+            int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart);
+            CHECK(poolSize);
             tgtData[(pd * outputH + ph) * outputW + pw] /= poolSize;
           }
         }
       }
       // compute offset
-      inData += inDepth * inHeight * inWidth;
-      tgtData += outputD * outputH * outputW;
+      inData += inLength;
+      tgtData += outLength;
     }
   }
 }
@@ -2437,8 +2411,10 @@ void CpuMatrix::avgPool3DBackward(Matrix& input,
                                   real scaleTargets,
                                   real scaleOutput) {
   size_t num = input.getHeight();
-  size_t channels = input.getWidth() / outputD / outputH / outputW;
-  CHECK(imgSizeD * imgSizeH * imgSizeW * channels == getWidth());
+  size_t inLength = imgSizeH * imgSizeW * imgSizeD;
+  size_t outLength = outputH * outputW * outputD;
+  size_t channels = input.getWidth() / outLength;
+  CHECK(inLength * channels == getWidth());
   real* inData = input.getData();
   real* outData = getData();
 
@@ -2448,21 +2424,18 @@ void CpuMatrix::avgPool3DBackward(Matrix& input,
     }
     for (size_t c = 0; c < channels; ++c) {
       for (size_t pd = 0; pd < outputD; ++pd) {
+        int dstart = pd * strideD - paddingD;
+        int dend = std::min(dstart + sizeZ, imgSizeD);
+        dstart = std::max(dstart, 0);
         for (size_t ph = 0; ph < outputH; ++ph) {
+          int hstart = ph * strideH - paddingH;
+          int hend = std::min(hstart + sizeY, imgSizeH);
+          hstart = std::max(hstart, 0);
           for (size_t pw = 0; pw < outputW; ++pw) {
-            int dstart = pd * strideD - paddingD;
-            int hstart = ph * strideH - paddingH;
             int wstart = pw * strideW - paddingW;
-            int dend = std::min(dstart + sizeZ, imgSizeD + paddingD);
-            int hend = std::min(hstart + sizeY, imgSizeH + paddingH);
-            int wend = std::min(wstart + sizeX, imgSizeW + paddingW);
-            int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart);
-            dstart = std::max(dstart, 0);
-            hstart = std::max(hstart, 0);
+            int wend = std::min(wstart + sizeX, imgSizeW);
             wstart = std::max(wstart, 0);
-            dend = std::min(dend, static_cast<int>(imgSizeD));
-            hend = std::min(hend, static_cast<int>(imgSizeH));
-            wend = std::min(wend, static_cast<int>(imgSizeW));
+            int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart);
             CHECK(poolSize);
             for (int d = dstart; d < dend; ++d) {
               for (int h = hstart; h < hend; ++h) {
@@ -2476,8 +2449,8 @@ void CpuMatrix::avgPool3DBackward(Matrix& input,
         }
       }
       // offset
-      outData += imgSizeD * imgSizeH * imgSizeW;
-      inData += outputD * outputH * outputW;
+      outData += inLength;
+      inData += outLength;
     }
   }
 }
diff --git a/paddle/math/NEONFunctions.cpp b/paddle/math/NEONFunctions.cpp
new file mode 100644
index 0000000000..3bf47901f1
--- /dev/null
+++ b/paddle/math/NEONFunctions.cpp
@@ -0,0 +1,55 @@
+/* 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 defined(__ARM_NEON__) || defined(__ARM_NEON)
+
+#include "NEONFunctions.h"
+#include <arm_neon.h>
+
+namespace paddle {
+namespace neon {
+
+// b[i] = a[i] > 0.0f ? a[i] : 0.0f
+void relu(const float* a, float* b, int len) {
+  int offset = len % 16;
+  float32x4_t ma0, ma1, ma2, ma3;
+  float32x4_t mb0, mb1, mb2, mb3;
+
+  float32x4_t zero = vdupq_n_f32(0.f);
+  for (int k = 0; k < len / 16; k++, a += 16, b += 16) {
+    ma0 = vld1q_f32(a);
+    ma1 = vld1q_f32(a + 4);
+    ma2 = vld1q_f32(a + 8);
+    ma3 = vld1q_f32(a + 12);
+
+    mb0 = vmaxq_f32(ma0, zero);
+    mb1 = vmaxq_f32(ma1, zero);
+    mb2 = vmaxq_f32(ma2, zero);
+    mb3 = vmaxq_f32(ma3, zero);
+
+    vst1q_f32(b, mb0);
+    vst1q_f32(b + 4, mb1);
+    vst1q_f32(b + 8, mb2);
+    vst1q_f32(b + 12, mb3);
+  }
+
+  for (int i = 0; i < offset; i++) {
+    b[i] = a[i] > 0.0f ? a[i] : 0.0f;
+  }
+}
+
+}  // namespace neon
+}  // namespace paddle
+
+#endif
diff --git a/paddle/math/NEONFunctions.h b/paddle/math/NEONFunctions.h
new file mode 100644
index 0000000000..69085e3335
--- /dev/null
+++ b/paddle/math/NEONFunctions.h
@@ -0,0 +1,23 @@
+/* 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
+
+namespace paddle {
+namespace neon {
+
+void relu(const float* a, float* b, int len);
+
+}  // namespace neon
+}  // namespace paddle
diff --git a/paddle/math/tests/test_matrixCompare.cpp b/paddle/math/tests/test_matrixCompare.cpp
index 103f06acc5..061fb22e3f 100644
--- a/paddle/math/tests/test_matrixCompare.cpp
+++ b/paddle/math/tests/test_matrixCompare.cpp
@@ -825,9 +825,8 @@ void testMaxPoolFwdBwd(int numSamples,
                        int strideW,
                        int padH,
                        int padW) {
-  int outH = 0, outW = 0;
-  outH = (imgSizeH - ksizeH + 2 * padH + strideH - 1) / strideH + 1;
-  outW = (imgSizeW - ksizeW + 2 * padW + strideW - 1) / strideW + 1;
+  int outH = outputSize(imgSizeH, ksizeH, padH, strideH, true);
+  int outW = outputSize(imgSizeW, ksizeW, padW, strideW, true);
 
   int inWidth = imgSizeH * imgSizeW * channels;
   MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false);
@@ -927,9 +926,8 @@ void testAvgPoolFwdBwd(int numSamples,
                        int strideW,
                        int padH,
                        int padW) {
-  int outH = 0, outW = 0;
-  outH = (imgSizeH - ksizeH + 2 * padH + strideH - 1) / strideH + 1;
-  outW = (imgSizeW - ksizeW + 2 * padW + strideW - 1) / strideW + 1;
+  int outH = outputSize(imgSizeH, ksizeH, padH, strideH, true);
+  int outW = outputSize(imgSizeW, ksizeW, padW, strideW, true);
 
   int inWidth = imgSizeH * imgSizeW * channels;
   MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false);
diff --git a/paddle/memory/memcpy.cc b/paddle/memory/memcpy.cc
index a19a3e3675..19ec9ba9b2 100644
--- a/paddle/memory/memcpy.cc
+++ b/paddle/memory/memcpy.cc
@@ -62,6 +62,24 @@ void Copy<platform::GPUPlace, platform::GPUPlace>(platform::GPUPlace dst_place,
   }
 }
 
+template <>
+void Copy<platform::CPUPlace, platform::GPUPlace>(platform::CPUPlace dst_place,
+                                                  void* dst,
+                                                  platform::GPUPlace src_place,
+                                                  const void* src, size_t num) {
+  platform::SetDeviceId(src_place.device);
+  platform::GpuMemcpySync(dst, src, num, cudaMemcpyDeviceToHost);
+}
+
+template <>
+void Copy<platform::GPUPlace, platform::CPUPlace>(platform::GPUPlace dst_place,
+                                                  void* dst,
+                                                  platform::CPUPlace src_place,
+                                                  const void* src, size_t num) {
+  platform::SetDeviceId(dst_place.device);
+  platform::GpuMemcpySync(dst, src, num, cudaMemcpyHostToDevice);
+}
+
 #endif  // PADDLE_ONLY_CPU
 
 }  // namespace memory
diff --git a/paddle/operators/CMakeLists.txt b/paddle/operators/CMakeLists.txt
index f9ea25ab04..e3e934bccc 100644
--- a/paddle/operators/CMakeLists.txt
+++ b/paddle/operators/CMakeLists.txt
@@ -1,5 +1,7 @@
 file(GLOB GENERAL_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*_op.cc")
 string(REPLACE ".cc" "" GENERAL_OPS "${GENERAL_OPS}")
+set(pybind_file ${PADDLE_SOURCE_DIR}/paddle/pybind/pybind.h)
+file(WRITE ${pybind_file} "// Generated by the paddle/operator/CMakeLists.txt.  DO NOT EDIT!\n\n")
 function(op_library TARGET)
     # op_library is a function to create op library. The interface is same as
     # cc_library. But it handle split GPU/CPU code and link some common library
@@ -7,10 +9,11 @@ function(op_library TARGET)
     set(OP_LIBRARY ${TARGET} ${OP_LIBRARY} PARENT_SCOPE)
     set(cc_srcs)
     set(cu_srcs)
-    set(op_common_deps operator op_registry)
+    set(op_common_deps operator op_registry math_function)
     set(options "")
     set(oneValueArgs "")
     set(multiValueArgs SRCS DEPS)
+    set(pybind_flag 0)
     cmake_parse_arguments(op_library "${options}" "${oneValueArgs}"
             "${multiValueArgs}" ${ARGN})
 
@@ -46,22 +49,42 @@ function(op_library TARGET)
         cc_library(${TARGET} SRCS ${cc_srcs} DEPS ${op_library_DEPS}
                 ${op_common_deps})
     endif()
+
+    # net_op doesn't need pybind
+    if ("${TARGET}" STREQUAL "net_op")
+        set(pybind_flag 1)
+    endif()
+
+    # pybind USE_NO_KERNEL_OP
+    file(READ ${TARGET}.cc TARGET_CONTENT)
+    string(REGEX MATCH "OperatorWithKernel" regex_result "${TARGET_CONTENT}")
+    string(REPLACE "_op" "" TARGET "${TARGET}")
+    if (${pybind_flag} EQUAL 0 AND regex_result STREQUAL "")
+        file(APPEND ${pybind_file} "USE_NO_KERNEL_OP(${TARGET});\n")
+        set(pybind_flag 1)
+    endif()
+
+    # pybind USE_CPU_ONLY_OP
+    list(LENGTH cu_srcs cu_srcs_len)
+    if (${pybind_flag} EQUAL 0 AND ${cu_srcs_len} EQUAL 0)
+        file(APPEND ${pybind_file} "USE_CPU_ONLY_OP(${TARGET});\n")
+        set(pybind_flag 1)
+    endif()
+
+    # pybind USE_OP
+    if (${pybind_flag} EQUAL 0)
+        file(APPEND ${pybind_file} "USE_OP(${TARGET});\n")
+    endif()
 endfunction()
 
 add_subdirectory(math)
 
 set(DEPS_OPS
-    identity_op
-    minus_op
-    mul_op
     recurrent_op
-    scale_op)
-op_library(identity_op DEPS scale_op)
-op_library(minus_op DEPS scale_op)
-op_library(mul_op DEPS math_function)
+    cond_op)
 op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc
-  DEPS framework_proto tensor operator net_op)
-op_library(scale_op DEPS net_op)
+  DEPS framework_proto tensor net_op)
+op_library(cond_op SRCS cond_op.cc DEPS framework_proto tensor operator net_op)
 
 list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS})
 foreach(src ${GENERAL_OPS})
diff --git a/paddle/operators/accuracy_op.cc b/paddle/operators/accuracy_op.cc
new file mode 100644
index 0000000000..0c813748b2
--- /dev/null
+++ b/paddle/operators/accuracy_op.cc
@@ -0,0 +1,71 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/accuracy_op.h"
+
+namespace paddle {
+namespace operators {
+
+class AccuracyOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.InputVar("Inference"),
+        "Input(Inference) of AccuracyOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Label"),
+                            "Input(Label) of AccuracyOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Accuracy"),
+        "Output(Accuracy) of AccuracyOp should not be null.");
+
+    auto *inference = ctx.Input<framework::Tensor>("Inference");
+    auto *label = ctx.Input<framework::Tensor>("Label");
+
+    PADDLE_ENFORCE_EQ(label->dims().size(), 1, "label must be a vector");
+    PADDLE_ENFORCE_EQ(inference->dims()[0], label->dims()[0],
+                      "inference size must be the same as label size");
+
+    ctx.Output<framework::LoDTensor>("Accuracy")->Resize({1});
+  }
+};
+
+class AccuracyOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  AccuracyOpMaker(framework::OpProto *proto,
+                  framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    // TODO(typhoonzero): support both inference value and indices.
+    AddInput("Inference", "topk(indices) the network output");
+    AddInput("Label", "Label of the training data");
+    // TODO(typhoonzero): AddInput("Weight", ...
+    AddOutput("Accuracy", "The accuracy of current batch");
+
+    AddComment(
+        R"DOC(Accuracy. It will print accuracy rate for classification.
+The accuracy is:
+..  math::
+accuracy = \\frac{NumOfCorrectPredicts}{NumOfAllSamples})DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(accuracy, ops::AccuracyOp, ops::AccuracyOpMaker);
+REGISTER_OP_CPU_KERNEL(accuracy,
+                       ops::AccuracyKernel<paddle::platform::CPUPlace, float>);
diff --git a/paddle/operators/accuracy_op.cu b/paddle/operators/accuracy_op.cu
new file mode 100644
index 0000000000..0a6a0fd15c
--- /dev/null
+++ b/paddle/operators/accuracy_op.cu
@@ -0,0 +1,81 @@
+/* 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 <thrust/execution_policy.h>
+#include <thrust/reduce.h>
+#include "paddle/operators/accuracy_op.h"
+#include "paddle/platform/cuda_helper.h"
+
+namespace paddle {
+namespace operators {
+using platform::PADDLE_CUDA_NUM_THREADS;
+
+template <int BlockSize>
+__global__ void AccuracyCudaKernel(const int N, const int D, const int* Xdata,
+                                   const int* labeldata, float* accuracy) {
+  int count = 0;
+  __shared__ int total[BlockSize];
+
+  // support only 1 block
+  for (int i = threadIdx.x; i < (N); i += BlockSize) {
+    for (int j = 0; j < D; ++j) {
+      if (Xdata[i * D + j] == labeldata[i]) {
+        ++count;
+        break;
+      }
+    }
+  }
+  total[threadIdx.x] = count;
+  __syncthreads();
+
+  // reduce the count with init value 0, and output accuracy.
+  int result = thrust::reduce(thrust::device, total, total + BlockSize, 0);
+  if (threadIdx.x == 0) {
+    *accuracy = static_cast<float>(result) / static_cast<float>(N);
+  }
+}
+
+template <typename T>
+class AccuracyOpCUDAKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
+                   "It must use GPUPlace.");
+    auto* inference = ctx.Input<Tensor>("Inference");
+    auto* label = ctx.Input<Tensor>("Label");
+    auto* accuracy = ctx.Output<Tensor>("Accuracy");
+    // FIXME(typhoonzero): only support indices currently
+    // if add support for output values, how to detect the data type?
+    const int* inference_data = inference->data<int>();
+    const int* label_data = label->data<int>();
+    float* accuracy_data = accuracy->mutable_data<float>(ctx.GetPlace());
+
+    size_t num_samples = inference->dims()[0];
+    size_t infer_width = inference->dims()[1];
+    cudaMemset((void**)&accuracy_data, 0, sizeof(float));
+
+    if (num_samples == 0) {
+      return;
+    }
+
+    AccuracyCudaKernel<PADDLE_CUDA_NUM_THREADS><<<1, PADDLE_CUDA_NUM_THREADS>>>(
+        num_samples, infer_width, inference_data, label_data, accuracy_data);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OP_GPU_KERNEL(accuracy,
+                       paddle::operators::AccuracyOpCUDAKernel<float>);
diff --git a/paddle/operators/accuracy_op.h b/paddle/operators/accuracy_op.h
new file mode 100644
index 0000000000..fe704efe1c
--- /dev/null
+++ b/paddle/operators/accuracy_op.h
@@ -0,0 +1,77 @@
+/* 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 <algorithm>
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
+
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
+
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenScalar = framework::EigenScalar<T, MajorType, IndexType>;
+
+template <typename Place, typename T>
+class AccuracyKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* inference = ctx.Input<Tensor>("Inference");
+    auto* label = ctx.Input<Tensor>("Label");
+    auto* accuracy = ctx.Output<Tensor>("Accuracy");
+
+    float* accuracy_data = accuracy->mutable_data<float>(ctx.GetPlace());
+
+    const T* inference_data = inference->data<T>();
+    const T* label_data = label->data<T>();
+
+    size_t num_samples = inference->dims()[0];
+    size_t class_dim = inference->dims()[1];
+    *accuracy_data = 0.0f;
+
+    if (num_samples == 0) {
+      return;
+    }
+
+    int num_correct = 0;
+    // assume inference is already the topk of the output
+    for (size_t i = 0; i < num_samples; ++i) {
+      PADDLE_ENFORCE_GE(label_data[i], 0, "label must >= 0");
+      for (size_t j = 0; j < class_dim; ++j) {
+        if (inference_data[i * class_dim + j] == label_data[i]) {
+          ++num_correct;
+          break;
+        }
+      }
+    }
+
+    // FIXME(typhoonzero): we don't accumulate the accuracy for now.
+    *accuracy_data =
+        static_cast<float>(num_correct) / static_cast<float>(num_samples);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/operators/add_op.cc b/paddle/operators/add_op.cc
index 8dbd47cf0d..e83c1efeaf 100644
--- a/paddle/operators/add_op.cc
+++ b/paddle/operators/add_op.cc
@@ -23,10 +23,18 @@ class AddOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of AddOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
+                            "Input(Y) of AddOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of AddOp should not be null.");
+
     PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("X")->dims(),
                       ctx.Input<Tensor>("Y")->dims(),
                       "Two input of Add Op's dimension must be same.");
-    ctx.Output<Tensor>("Out")->Resize(ctx.Input<Tensor>("X")->dims());
+    ctx.Output<framework::LoDTensor>("Out")->Resize(
+        ctx.Input<Tensor>("X")->dims());
   }
 };
 
diff --git a/paddle/operators/concat_op.cc b/paddle/operators/concat_op.cc
new file mode 100644
index 0000000000..223bb0ffe6
--- /dev/null
+++ b/paddle/operators/concat_op.cc
@@ -0,0 +1,82 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/concat_op.h"
+#include <vector>
+
+namespace paddle {
+namespace operators {
+using framework::Tensor;
+
+class ConcatOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of ConcatOp should not be null.");
+
+    auto ins = ctx.MultiInput<framework::Tensor>("X");
+    auto *out = ctx.Output<framework::LoDTensor>("Out");
+    size_t axis = static_cast<size_t>(ctx.Attr<int>("axis"));
+    size_t n = ins.size();
+
+    PADDLE_ENFORCE_GT(n, 1, "Input tensors count should > 1.");
+
+    auto out_dims = ins[0]->dims();
+    size_t in_zero_dims_size = out_dims.size();
+    for (size_t i = 1; i < n; i++) {
+      for (size_t j = 0; j < in_zero_dims_size; j++) {
+        if (j == axis) {
+          out_dims[axis] += ins[i]->dims()[j];
+          continue;
+        }
+        PADDLE_ENFORCE_EQ(out_dims[j], ins[i]->dims()[j],
+                          "Input tensors should have the same "
+                          "elements except the specify axis.")
+      }
+    }
+    out->Resize(out_dims);
+  }
+};
+
+class ConcatOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  ConcatOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "the input tensors of concat operator.").AsDuplicable();
+    AddOutput("Out", "the output tensor of concat operator.");
+    AddComment(R"DOC(
+            Join the input tensors along with the axis.
+            Examples:
+              Input[0] = [[1,2],[3,4]]
+              Input[1] = [[5,6]]
+              axis = 0
+              Output = [[1,2],
+                        [3,4],
+                        [5,6]]
+        )DOC");
+    AddAttr<int>("axis", "The axis which the inputs will be joined with.")
+        .SetDefault(0);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(concat, ops::ConcatOp, ops::ConcatOpMaker)
+REGISTER_OP_CPU_KERNEL(concat,
+                       ops::ConcatKernel<paddle::platform::CPUPlace, float>)
diff --git a/paddle/operators/concat_op.h b/paddle/operators/concat_op.h
new file mode 100644
index 0000000000..f977054fdf
--- /dev/null
+++ b/paddle/operators/concat_op.h
@@ -0,0 +1,64 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include <vector>
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename Place, typename T>
+class ConcatKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto ins = ctx.MultiInput<framework::Tensor>("X");
+    auto* out = ctx.Output<framework::Tensor>("Out");
+    int64_t axis = static_cast<int64_t>(ctx.Attr<int>("axis"));
+    size_t n = ins.size();
+    size_t output_axis_dim = 0;
+    size_t before = 1, after = 1;
+    for (size_t i = 0; i < n; i++) {
+      output_axis_dim += ins[i]->dims()[axis];
+    }
+    auto& input_zero = ins[0];
+    for (int64_t i = 0; i < input_zero->dims().size(); i++) {
+      if (i == axis) {
+        continue;
+      }
+      if (i < axis) {
+        before *= input_zero->dims()[i];
+      } else {
+        after *= input_zero->dims()[i];
+      }
+    }
+    size_t output_offset = 0;
+    for (size_t i = 0; i < n; i++) {
+      auto& in = ins[i];
+      auto axis_dim = in->dims()[axis];
+      for (size_t j = 0; j < before; j++) {
+        size_t len = axis_dim * after * sizeof(T);
+        const T* src = in->data<T>() + axis_dim * after * j;
+        T* out_data = out->mutable_data<T>(platform::CPUPlace());
+        T* dest = out_data + output_offset + output_axis_dim * after * j;
+        memcpy(dest, src, len);
+      }
+      output_offset += axis_dim * after;
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/operators/cond_op.cc b/paddle/operators/cond_op.cc
new file mode 100644
index 0000000000..8262a7a5c8
--- /dev/null
+++ b/paddle/operators/cond_op.cc
@@ -0,0 +1,229 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/cond_op.h"
+
+#include <cstring>
+#include <sstream>
+
+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/gather.h"
+#include "paddle/operators/net_op.h"
+#include "paddle/operators/scatter.h"
+
+namespace paddle {
+namespace operators {
+
+using Scope = framework::Scope;
+using Variable = framework::Variable;
+using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
+using DDim = framework::DDim;
+
+void CondOp::CreateScope(const Scope& scope) const {
+  auto sub_scopes_var = scope.FindVar("SubScopes");
+  PADDLE_ENFORCE_NOT_NULL(sub_scopes_var,
+                          "Output(SubScopes) of CondOp should not be null.");
+  auto sub_scopes = sub_scopes_var->GetMutable<std::vector<Scope*>>();
+  auto& sub_scope = scope.NewScope();
+  sub_scopes->push_back(&sub_scope);
+}
+
+void CondOp::CreateIndexTensor(const Scope& scope) const {
+  auto index_tensors_var = scope.FindVar("IndexTensors");
+  PADDLE_ENFORCE_NOT_NULL(index_tensors_var,
+                          "Output(IndexTensors) of CondOp should not be null.");
+  auto& index_tensors =
+      *index_tensors_var->GetMutable<std::vector<LoDTensor>>();
+  index_tensors.push_back(LoDTensor());
+}
+
+void CondOp::InferShape(const Scope& scope) const {
+  auto sub_scopes_var = scope.FindVar("SubScopes");
+  PADDLE_ENFORCE_NOT_NULL(sub_scopes_var,
+                          "Output(SubScopes) of CondOp should not be null.");
+  auto& sub_scopes = *sub_scopes_var->GetMutable<std::vector<Scope*>>();
+
+  for (int i = 0; i < 2; ++i) {
+    // Create two sub scopes for true and false branches
+    // sub_scopes[0] for the true branch and sub_scopes[1] for the false
+    // branch
+    CreateScope(scope);
+
+    // Create two tensors for true and false indices
+    // index_tensors[0] for the true branch and index_tensors[1] for the false
+    // branch
+    CreateIndexTensor(scope);
+
+    PADDLE_ENFORCE(!Inputs("Xs").empty(),
+                   "Inputs(Xs) of CondOp can't be empty.");
+    for (auto& input : Inputs("Xs")) {
+      // Create a new tensor in sub-scope for input-type tensor
+      Variable* v = sub_scopes[i]->NewVar(input);
+      LoDTensor* sub_input = v->GetMutable<LoDTensor>();
+      sub_input->Resize(scope.FindVar(input)->GetMutable<LoDTensor>()->dims());
+    }
+
+    for (auto& output : (*sub_net_op_[i]).Outputs()) {
+      for (auto& var_name : output.second) {
+        sub_scopes[i]->NewVar(var_name);
+      }
+    }
+
+    // each net calls InferShape
+    sub_net_op_[i]->InferShape(*sub_scopes[i]);
+  }
+
+  for (auto& output : Outputs("Outs")) {
+    LoDTensor* tensor_t_out =
+        sub_scopes[0]->FindVar(output)->GetMutable<LoDTensor>();
+    PADDLE_ENFORCE_NOT_NULL(tensor_t_out, "True output should not be NULL");
+    LoDTensor* tensor_f_out =
+        sub_scopes[1]->FindVar(output)->GetMutable<LoDTensor>();
+    PADDLE_ENFORCE_NOT_NULL(tensor_f_out, "False output should not be NULL");
+
+    auto* tensor_out_var = scope.FindVar(output);
+    PADDLE_ENFORCE_NOT_NULL(tensor_out_var, "Output not found");
+    LoDTensor* tensor_out = tensor_out_var->GetMutable<LoDTensor>();
+    PADDLE_ENFORCE_NOT_NULL(tensor_t_out,
+                            "True output tensor should not be NULL");
+
+    // check output size should be same
+    PADDLE_ENFORCE_EQ(tensor_t_out->dims(), tensor_f_out->dims(),
+                      "Outputs not of the same shape");
+    tensor_out->Resize(tensor_t_out->dims());
+    // tensor_out->mutable_data<float>(tensor_out->dims(),
+    // platform::CPUPlace());
+    tensor_out->mutable_data<float>(platform::CPUPlace());
+  }
+}
+
+void CondOp::Run(const Scope& scope,
+                 const platform::DeviceContext& dev_ctx) const {
+  auto* sub_scopes_var = scope.FindVar("SubScopes");
+  PADDLE_ENFORCE_NOT_NULL(sub_scopes_var,
+                          "Output(SubScopes) of CondOp should not be null.");
+  auto sub_scopes = sub_scopes_var->Get<std::vector<Scope*>>();
+  auto* index_tensors_var = scope.FindVar("IndexTensors");
+  PADDLE_ENFORCE_NOT_NULL(index_tensors_var,
+                          "Output(IndexTensors) of CondOp should not be null.");
+  auto index_tensors = index_tensors_var->Get<std::vector<LoDTensor>>();
+
+  std::string cond_name = Input("Cond");
+  Variable* cond_var = scope.FindVar(cond_name);
+  PADDLE_ENFORCE_NOT_NULL(cond_var,
+                          "Input(Cond) of CondOp should not be null.");
+  const LoDTensor* cond = cond_var->GetMutable<LoDTensor>();
+
+  // Step 1: get the true/false index at runtime
+  // index_[0]: vector<int>, contains all index for cond[i] == true
+  // index_[1]: vector<int>, contains all index for cond[i] == false
+  for (int i = 0; i < 2; ++i) index_[i].clear();
+
+  const int* cond_data = cond->data<int>();
+  for (int i = 0; i < cond->dims()[0]; ++i) {
+    if (cond_data[i])
+      index_[0].push_back(i);
+    else
+      index_[1].push_back(i);
+  }
+
+  // put index_[0] and index_[1] into two tensors:
+  // index_tensor_[0] and index_tensor_[1]
+  DDim dim = paddle::framework::make_ddim({0});
+  for (int i = 0; i < 2; ++i) {
+    dim[0] = index_[i].size();
+    int* tmp_ptr =
+        index_tensors[i].mutable_data<int>(dim, platform::CPUPlace());
+    index_tensors[i].Resize(dim);
+    memcpy(tmp_ptr, index_[i].data(), dim[0] * sizeof(int));
+  }
+
+  // Step 2: collect data by calling gather
+  for (int i = 0; i < 2; ++i) {
+    // i= 0/i for True and False branches respectively
+    for (auto& input : Inputs("Xs")) {
+      // find Tensor
+      Variable* v = scope.FindVar(input);
+      PADDLE_ENFORCE_NOT_NULL(v);
+      LoDTensor* tensor_parent = v->GetMutable<LoDTensor>();
+
+      v = sub_scopes[i]->FindVar(input);
+      PADDLE_ENFORCE_NOT_NULL(v);
+      LoDTensor* tensor_child = v->GetMutable<LoDTensor>();
+
+      // Resize child
+      DDim dim = tensor_child->dims();
+      dim[0] = index_[i].size();
+      tensor_child->Resize(dim);
+      tensor_child->mutable_data<float>(dim, platform::CPUPlace());
+
+      Gather<float>(dev_ctx.GetPlace(), tensor_parent, &index_tensors[i],
+                    tensor_child);
+    }
+  }
+
+  // Step 3: run
+  for (int i = 0; i < 2; ++i) {
+    sub_net_op_[i]->Run(*sub_scopes[i], dev_ctx);
+  }
+
+  // Step 4: merge output results
+  PADDLE_ENFORCE(!Outputs("Outs").empty(),
+                 "Outputs(Outs) of CondOp can't be empty.");
+  for (int i = 0; i < 2; ++i) {
+    // i= 0/i for True and False branches respectively
+    for (auto& output : Outputs("Outs")) {
+      // find Tensor
+      Variable* v = scope.FindVar(output);
+      PADDLE_ENFORCE_NOT_NULL(v);
+      LoDTensor* tensor_parent = v->GetMutable<LoDTensor>();
+
+      v = sub_scopes[i]->FindVar(output);
+      PADDLE_ENFORCE_NOT_NULL(v);
+      LoDTensor* tensor_child = v->GetMutable<LoDTensor>();
+
+      ScatterUpdate<float>(dev_ctx.GetPlace(), tensor_child, &index_tensors[i],
+                           tensor_parent);
+    }
+  }
+}
+
+class CondOpProtoAndCheckerMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  CondOpProtoAndCheckerMaker(framework::OpProto* proto,
+                             framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("Cond", "The condition, which is a bool vector");
+    AddInput("Xs", "Inputs of Subnets").AsDuplicable();
+    AddOutput("Outs", "Outputs of Cond_Op after merge").AsDuplicable();
+
+    AddOutput("SubScopes", "sub scopes for true and false branches");
+    AddOutput("IndexTensors", "Index Tensors contains indices for true/false");
+
+    AddComment(R"DOC(
+Sample dependent Cond Operator:
+Given Cond[i] as a 1/0 vector to indicate true/false
+The equation is: 
+Out[i] = subnet_t[i], if Cond[i] == true
+Out[i] = subnet_t[i], if Cond[i] == false
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OP_WITHOUT_GRADIENT(cond, paddle::operators::CondOp,
+                             paddle::operators::CondOpProtoAndCheckerMaker);
diff --git a/paddle/operators/cond_op.h b/paddle/operators/cond_op.h
new file mode 100644
index 0000000000..b09e32331e
--- /dev/null
+++ b/paddle/operators/cond_op.h
@@ -0,0 +1,91 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include <vector>
+#include "glog/logging.h"
+#include "paddle/framework/ddim.h"
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/operator.h"
+#include "paddle/framework/tensor.h"
+#include "paddle/operators/net_op.h"
+
+namespace paddle {
+namespace operators {
+
+/*
+ * @brief CondOp is a dynamic if-else Operator
+ *
+ * It has a input tensor named cond indicating which netop each instance will
+ * run.
+ *
+ * if cond == 1, it will run true_net, which is a NetOp.
+ *
+ * if cond == 0, it will run false_net, which is another NetOp.
+ */
+class CondOp : public framework::OperatorBase {
+ public:
+  CondOp(const std::string& type, const framework::VariableNameMap& inputs,
+         const framework::VariableNameMap& outputs,
+         const framework::AttributeMap& attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {
+    index_.resize(2);
+    sub_net_op_.resize(2);
+  }
+
+  CondOp(const CondOp& o)
+      : framework::OperatorBase(
+            static_cast<const framework::OperatorBase&>(o)) {
+    // TODO(yuyang18): Implement copy ctor well.
+    PADDLE_THROW("Not implemented");
+  }
+
+  void CreateScope(const framework::Scope& scope) const;
+
+  void CreateIndexTensor(const framework::Scope& scope) const;
+
+  /*
+   * InferShape must be called before Run.
+   */
+  void InferShape(const framework::Scope& scope) const override;
+
+  /*
+   * Set True Block
+   */
+  void set_truenet(std::unique_ptr<OperatorBase>&& net) {
+    sub_net_op_[0] = std::move(net);
+  }
+
+  /*
+   * Set False Block
+   */
+  void set_falsenet(std::unique_ptr<OperatorBase>&& net) {
+    sub_net_op_[1] = std::move(net);
+  }
+
+  void Run(const framework::Scope& scope,
+           const platform::DeviceContext& dev_ctx) const override;
+
+ private:
+  // sub_net_op_[0]: subnet_t
+  // sub_net_op_[1]: subnet_f
+  std::vector<std::unique_ptr<framework::OperatorBase>> sub_net_op_;
+
+  // index_[0]: True_index;
+  // index_[1]: False_index;
+  mutable std::vector<std::vector<int>> index_;
+};
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/operators/cos_sim_op.cc b/paddle/operators/cos_sim_op.cc
index c033af3b74..72c4464936 100644
--- a/paddle/operators/cos_sim_op.cc
+++ b/paddle/operators/cos_sim_op.cc
@@ -25,16 +25,38 @@ class CosSimOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) must not be null.");
-    PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("X")->dims(),
-                      ctx.Input<Tensor>("Y")->dims(),
-                      "Dimensions of Input(X) and Input(Y) must be the same.");
-
-    auto dims = ctx.Input<Tensor>("X")->dims();
-    ctx.Output<Tensor>("Out")->Resize({dims[0], 1});
-    ctx.Output<Tensor>("XNorm")->Resize({dims[0], 1});
-    ctx.Output<Tensor>("YNorm")->Resize({dims[0], 1});
+    // notnull check
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of CosSimOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
+                            "Input(Y) of CosSimOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of CosSimOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("XNorm"),
+                            "Output(XNorm) of CosSimOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("YNorm"),
+                            "Output(YNorm) of CosSimOp should not be null.");
+
+    // shape check
+    auto x_dims = ctx.Input<Tensor>("X")->dims();
+    auto y_dims = ctx.Input<Tensor>("Y")->dims();
+
+    PADDLE_ENFORCE_EQ(x_dims.size(), y_dims.size(),
+                      "Ranks of Input(X) and Input(Y) must be equal.");
+    PADDLE_ENFORCE_GE(x_dims.size(), 2,
+                      "Rank of Input(X) must not be less than 2.");
+    PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 1, x_dims.size()),
+                      framework::slice_ddim(y_dims, 1, y_dims.size()),
+                      "All dimensions except the 1st of Input(X) and Input(Y) "
+                      "must be equal.");
+    PADDLE_ENFORCE(x_dims[0] == y_dims[0] || y_dims[0] == 1,
+                   "The 1st dimension of Input(Y) must be equal to Input(X) or"
+                   " just 1 (which will be broadcasted to match Input(X)).");
+
+    // resize tensor
+    ctx.Output<framework::LoDTensor>("Out")->Resize({x_dims[0], 1});
+    ctx.Output<framework::LoDTensor>("XNorm")->Resize({x_dims[0], 1});
+    ctx.Output<framework::LoDTensor>("YNorm")->Resize({y_dims[0], 1});
   }
 };
 
@@ -42,16 +64,27 @@ class CosSimOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   CosSimOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "The first input of cos_sim op.");
-    AddInput("Y", "The second input of cos_sim op.");
+    AddInput("X", "The 1st input of cos_sim op.");
+    AddInput("Y", "The 2nd input of cos_sim op.");
     AddOutput("Out", "The output of cos_sim op.");
-    AddOutput("XNorm", "Row norm of the first input.").AsIntermediate();
-    AddOutput("YNorm", "Row norm of the second input.").AsIntermediate();
+    AddOutput("XNorm",
+              "Norm of the first input, reduced along the 1st "
+              "dimension.")
+        .AsIntermediate();
+    AddOutput("YNorm",
+              "Norm of the second input, reduced along the 1st "
+              "dimension.")
+        .AsIntermediate();
 
     AddComment(R"DOC(
 Cosine Similarity Operator.
 
-The equation is: Out = X^T * Y / (sqrt(X^T * X) * sqrt(Y^T * Y))
+The equation is: Out = X^T * Y / (sqrt(X^T * X) * sqrt(Y^T * Y)).
+
+Input(X) and Input(Y) must have the same shape, except that the 1st dimension
+of Input(Y) could be just 1 (different from Input(X)), which will be
+broadcasted to match the shape of Input(X) before computing their cosine
+similarity.
 )DOC");
   }
 };
@@ -62,34 +95,54 @@ class CosSimOpGrad : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    // notnull check
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null.");
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) must not be null.");
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("XNorm"),
                             "Input(XNorm) must not be null.");
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("YNorm"),
                             "Input(YNorm) must not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Out"),
+                            "Input(Out) must not be null.");
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
                             "Input(Out@GRAD) must not be null.");
 
+    // shape check
     auto x_dims = ctx.Input<Tensor>("X")->dims();
     auto y_dims = ctx.Input<Tensor>("Y")->dims();
     auto xnorm_dims = ctx.Input<Tensor>("XNorm")->dims();
     auto ynorm_dims = ctx.Input<Tensor>("YNorm")->dims();
-    auto out_dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
-    PADDLE_ENFORCE_EQ(x_dims, y_dims,
-                      "Dimensions of Input(X) and Input(Y) must be the same.");
-    PADDLE_ENFORCE_EQ(xnorm_dims[0], x_dims[0],
-                      "1st dimension of XNorm must equal that of Input(X).");
-    PADDLE_ENFORCE_EQ(xnorm_dims[1], 1, "2st dimension of XNorm must be one.");
-    PADDLE_ENFORCE_EQ(ynorm_dims[0], y_dims[0],
-                      "1st dimension of YNorm must equal that of Input(Y).");
-    PADDLE_ENFORCE_EQ(ynorm_dims[1], 1, "2st dimension of YNorm must be one.");
-    PADDLE_ENFORCE_EQ(out_dims[0], x_dims[0],
-                      "1st dimension of Out@GRAD must equal that of Input(X)");
-    PADDLE_ENFORCE_EQ(out_dims[1], 1, "1st dimension of Out@GRAD must be one.");
-
-    auto *x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
-    auto *y_grad = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    auto out_dims = ctx.Input<Tensor>("Out")->dims();
+    auto out_grad_dims =
+        ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
+
+    PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
+                      "Ranks of Input(X) and Input(Y) must be equal.");
+    PADDLE_ENFORCE_GE(x_dims.size(), 2,
+                      "Rank of Input(X) must not be less than 2.");
+    PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 1, x_dims.size()),
+                      framework::slice_ddim(y_dims, 1, y_dims.size()),
+                      "All dimensions except the 1st of Input(X) and Input(Y) "
+                      "must be equal.");
+    PADDLE_ENFORCE(x_dims[0] == y_dims[0] || y_dims[0] == 1,
+                   "The 1st dimension of Input(Y) must be equal to Input(X) or"
+                   " just 1 (which will be broadcasted to match Input(X)).");
+    auto target_xnorm_dims = framework::make_ddim({x_dims[0], 1});
+    auto target_ynorm_dims = framework::make_ddim({y_dims[0], 1});
+    PADDLE_ENFORCE_EQ(xnorm_dims, target_xnorm_dims,
+                      "Shape of Input(XNorm) must be [X.Dim(0), 1].");
+    PADDLE_ENFORCE_EQ(ynorm_dims, target_ynorm_dims,
+                      "Shape of Input(YNorm) must be [Y.Dim(0), 1].");
+    PADDLE_ENFORCE_EQ(out_dims, target_xnorm_dims,
+                      "Shape of Input(Out) must be [X.Dim(0), 1].");
+    PADDLE_ENFORCE_EQ(out_grad_dims, target_xnorm_dims,
+                      "Shape of Input(Out@Grad) must be [X.Dim(0), 1].");
+
+    // resize tensor
+    auto *x_grad =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto *y_grad =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("Y"));
     if (x_grad) x_grad->Resize(x_dims);
     if (y_grad) y_grad->Resize(y_dims);
   }
diff --git a/paddle/operators/cos_sim_op.h b/paddle/operators/cos_sim_op.h
index 9e2bcebe3b..bcf6f758ca 100644
--- a/paddle/operators/cos_sim_op.h
+++ b/paddle/operators/cos_sim_op.h
@@ -31,30 +31,38 @@ template <typename Place, typename T>
 class CosSimKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto* input_x = context.Input<Tensor>("X");
-    auto* input_y = context.Input<Tensor>("Y");
-    auto* output_z = context.Output<Tensor>("Out");
-    auto* output_x_norm = context.Output<Tensor>("XNorm");
-    auto* output_y_norm = context.Output<Tensor>("YNorm");
+    // get Tensor
+    auto* in_x = context.Input<Tensor>("X");
+    auto* in_y = context.Input<Tensor>("Y");
+    auto* out_z = context.Output<Tensor>("Out");
+    auto* out_x_norm = context.Output<Tensor>("XNorm");
+    auto* out_y_norm = context.Output<Tensor>("YNorm");
+    out_z->mutable_data<T>(context.GetPlace());
+    out_x_norm->mutable_data<T>(context.GetPlace());
+    out_y_norm->mutable_data<T>(context.GetPlace());
 
-    output_z->mutable_data<T>(context.GetPlace());
-    output_x_norm->mutable_data<T>(context.GetPlace());
-    output_y_norm->mutable_data<T>(context.GetPlace());
-
-    auto dims = input_x->dims();
-    int size = static_cast<int>(framework::product(dims));
-    auto new_dims = framework::make_ddim({dims[0], size / dims[0]});
-    auto x = EigenMatrix<T>::From(*input_x, new_dims);
-    auto y = EigenMatrix<T>::From(*input_y, new_dims);
-    auto z = EigenVector<T>::Flatten(*output_z);
-    auto x_norm = EigenVector<T>::Flatten(*output_x_norm);
-    auto y_norm = EigenVector<T>::Flatten(*output_y_norm);
+    // convert Tensor to Eigen Tensor
+    int rows_x = in_x->dims()[0];
+    int rows_y = in_y->dims()[0];
+    auto x = EigenMatrix<T>::Reshape(*in_x, 1);
+    auto y = EigenMatrix<T>::Reshape(*in_y, 1);
+    auto z = EigenVector<T>::Flatten(*out_z);
+    auto x_norm = EigenVector<T>::Flatten(*out_x_norm);
+    auto y_norm = EigenVector<T>::Flatten(*out_y_norm);
 
+    // compute
     auto place = context.GetEigenDevice<Place>();
-    auto xy = (x * y).sum(Eigen::array<int, 1>({{1}}));
-    x_norm.device(place) = x.square().sum(Eigen::array<int, 1>({{1}})).sqrt();
-    y_norm.device(place) = y.square().sum(Eigen::array<int, 1>({{1}})).sqrt();
-    z.device(place) = xy / x_norm / y_norm;
+    auto row_along = Eigen::array<int, 1>({{1}});
+    x_norm.device(place) = x.square().sum(row_along).sqrt();
+    y_norm.device(place) = y.square().sum(row_along).sqrt();
+    if (rows_x == rows_y) {
+      auto xy = (x * y).sum(Eigen::array<int, 1>({{1}}));
+      z.device(place) = xy / x_norm / y_norm;
+    } else {
+      Eigen::DSizes<int, 2> bcast(rows_x, 1);
+      auto xy = (x * y.broadcast(bcast)).sum(row_along);
+      z.device(place) = xy / x_norm / y_norm.broadcast(bcast);
+    }
   }
 };
 
@@ -62,43 +70,72 @@ template <typename Place, typename T>
 class CosSimGradKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto* input_x = context.Input<Tensor>("X");
-    auto* input_y = context.Input<Tensor>("Y");
-    auto* input_z = context.Input<Tensor>("Out");
-    auto* input_x_norm = context.Input<Tensor>("XNorm");
-    auto* input_y_norm = context.Input<Tensor>("YNorm");
-    auto* output_grad_x = context.Output<Tensor>(framework::GradVarName("X"));
-    auto* output_grad_y = context.Output<Tensor>(framework::GradVarName("Y"));
-    auto* input_grad_z = context.Input<Tensor>(framework::GradVarName("Out"));
+    // get Tensor
+    auto* in_x = context.Input<Tensor>("X");
+    auto* in_y = context.Input<Tensor>("Y");
+    auto* in_z = context.Input<Tensor>("Out");
+    auto* in_x_norm = context.Input<Tensor>("XNorm");
+    auto* in_y_norm = context.Input<Tensor>("YNorm");
+    auto* out_grad_x = context.Output<Tensor>(framework::GradVarName("X"));
+    auto* out_grad_y = context.Output<Tensor>(framework::GradVarName("Y"));
+    auto* in_grad_z = context.Input<Tensor>(framework::GradVarName("Out"));
 
-    auto dims = input_x->dims();
-    int size = static_cast<int>(framework::product(dims));
-    auto new_dims = framework::make_ddim({dims[0], size / dims[0]});
-    auto x = EigenMatrix<T>::From(*input_x, new_dims);
-    auto y = EigenMatrix<T>::From(*input_y, new_dims);
-    auto z = EigenMatrix<T>::From(*input_z);
-    auto x_norm = EigenMatrix<T>::From(*input_x_norm);
-    auto y_norm = EigenMatrix<T>::From(*input_y_norm);
-    auto dz = EigenMatrix<T>::From(*input_grad_z);
+    // convert Tensor to Eigen Tensor
+    auto x = EigenMatrix<T>::Reshape(*in_x, 1);
+    auto y = EigenMatrix<T>::Reshape(*in_y, 1);
+    auto z = EigenMatrix<T>::Reshape(*in_z, 1);
+    auto x_norm = EigenMatrix<T>::Reshape(*in_x_norm, 1);
+    auto y_norm = EigenMatrix<T>::Reshape(*in_y_norm, 1);
+    auto dz = EigenMatrix<T>::Reshape(*in_grad_z, 1);
 
-    Eigen::DSizes<int, 2> bcast(1, new_dims[1]);
-    auto z_bcast = z.broadcast(bcast);
-    auto dz_bcast = dz.broadcast(bcast);
+    // compute gradident
+    int rows_x = in_x->dims()[0];
+    int rows_y = in_y->dims()[0];
+    int cols = framework::product(in_x->dims()) / rows_x;
+    Eigen::DSizes<int, 2> bcast_cols(1, cols);
+    auto z_bcast = z.broadcast(bcast_cols);
+    auto dz_bcast = dz.broadcast(bcast_cols);
+    auto x_snorm_bcast = x_norm.square().eval().broadcast(bcast_cols);
     auto place = context.GetEigenDevice<Place>();
-    auto x_snorm_bcast = x_norm.square().eval().broadcast(bcast);
-    auto y_snorm_bcast = y_norm.square().eval().broadcast(bcast);
-    auto norm_prod_bcast = (x_norm * y_norm).eval().broadcast(bcast);
-    if (output_grad_x) {
-      output_grad_x->mutable_data<T>(context.GetPlace());
-      auto dx = EigenMatrix<T>::From(*output_grad_x, new_dims);
-      dx.device(place) =
-          dz_bcast * (y / norm_prod_bcast - z_bcast * x / x_snorm_bcast);
-    }
-    if (output_grad_y) {
-      output_grad_y->mutable_data<T>(context.GetPlace());
-      auto dy = EigenMatrix<T>::From(*output_grad_y, new_dims);
-      dy.device(place) =
-          dz_bcast * (x / norm_prod_bcast - z_bcast * y / y_snorm_bcast);
+    if (rows_x == rows_y) {
+      auto y_snorm_bcast = y_norm.square().eval().broadcast(bcast_cols);
+      auto norm_prod_bcast = (x_norm * y_norm).eval().broadcast(bcast_cols);
+      // compute dx
+      if (out_grad_x) {
+        out_grad_x->mutable_data<T>(context.GetPlace());
+        auto dx = EigenMatrix<T>::Reshape(*out_grad_x, 1);
+        auto grad = y / norm_prod_bcast - z_bcast * x / x_snorm_bcast;
+        dx.device(place) = dz_bcast * grad;
+      }
+      // compute dy
+      if (out_grad_y) {
+        out_grad_y->mutable_data<T>(context.GetPlace());
+        auto dy = EigenMatrix<T>::Reshape(*out_grad_y, 1);
+        auto grad = x / norm_prod_bcast - z_bcast * y / y_snorm_bcast;
+        dy.device(place) = dz_bcast * grad;
+      }
+    } else {
+      Eigen::DSizes<int, 2> bcast_rows(rows_x, 1);
+      Eigen::DSizes<int, 2> bcast_rows_cols(rows_x, cols);
+      auto y_bcast = y.broadcast(bcast_rows);
+      auto y_snorm_bcast = y_norm.square().eval().broadcast(bcast_rows_cols);
+      auto norm_prod_bcast = (x_norm * y_norm.eval().broadcast(bcast_rows))
+                                 .eval()
+                                 .broadcast(bcast_cols);
+      // compute dx
+      if (out_grad_x) {
+        out_grad_x->mutable_data<T>(context.GetPlace());
+        auto dx = EigenMatrix<T>::Reshape(*out_grad_x, 1);
+        auto grad = y_bcast / norm_prod_bcast - z_bcast * x / x_snorm_bcast;
+        dx.device(place) = dz_bcast * grad;
+      }
+      // compute dy
+      if (out_grad_y) {
+        out_grad_y->mutable_data<T>(context.GetPlace());
+        auto dy = EigenMatrix<T>::Reshape(*out_grad_y, 1);
+        auto grad = x / norm_prod_bcast - z_bcast * y_bcast / y_snorm_bcast;
+        dy.device(place) = (dz_bcast * grad).sum(Eigen::array<int, 1>({{0}}));
+      }
     }
   }
 };
diff --git a/paddle/operators/elementwise_mul_op.cc b/paddle/operators/elementwise_mul_op.cc
new file mode 100644
index 0000000000..ee6e975b44
--- /dev/null
+++ b/paddle/operators/elementwise_mul_op.cc
@@ -0,0 +1,117 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/operators/elementwise_mul_op.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+class ElementWiseMulOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of ElementWiseMulOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
+                            "Input(Y) of ElementWiseMulOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Out"),
+        "Output(Out) of ElementWiseMulOp should not be null.");
+
+    auto x_dim = ctx.Input<Tensor>("X")->dims();
+    auto y_dim = ctx.Input<Tensor>("Y")->dims();
+    PADDLE_ENFORCE_GE(x_dim.size(), y_dim.size(),
+                      "Rank of first input must >= rank of second input.")
+    ctx.Output<framework::LoDTensor>("Out")->Resize(x_dim);
+  }
+};
+
+class ElementWiseMulOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  ElementWiseMulOpMaker(framework::OpProto *proto,
+                        framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "The first input of elementwise mul op");
+    AddInput("Y", "The second input of elementwise mul op");
+    AddAttr<int>("axis",
+                 R"DOC(
+When shape(Y) does not equal shape(X),Y will be broadcasted 
+to match the shape of X and axis should be dimension index Y in X
+        )DOC")
+        .SetDefault(-1)
+        .EqualGreaterThan(-1);
+
+    AddOutput("Out", "The output of elementwise mul op");
+    AddComment(R"DOC(
+Limited elementwise multiple operator.The equation is: Out = X ⊙ Y.
+1. The shape of Y should be same with X or
+2. Y's shape is a subset of X. 
+   Y will be broadcasted to match the shape of X and axis should be dimension index Y in X.
+   example:
+      shape(X) = (2, 3, 4, 5), shape(Y) = (,)
+      shape(X) = (2, 3, 4, 5), shape(Y) = (5,)
+      shape(X) = (2, 3, 4, 5), shape(Y) = (4, 5)
+      shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1
+      shape(X) = (2, 3, 4, 5), shape(Y) = (2), with axis=0
+)DOC");
+  }
+};
+
+class ElementWiseMulOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should not be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
+                            "Input(Out@GRAD) should not be null");
+
+    auto x_dims = ctx.Input<Tensor>("X")->dims();
+    auto y_dims = ctx.Input<Tensor>("Y")->dims();
+    auto out_dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
+    auto *x_grad =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto *y_grad =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("Y"));
+
+    PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
+                      "Rank of first input must >= rank of second input.")
+
+    if (x_grad) {
+      x_grad->Resize(x_dims);
+    }
+
+    if (y_grad) {
+      y_grad->Resize(y_dims);
+    }
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(elementwise_mul, ops::ElementWiseMulOp, ops::ElementWiseMulOpMaker,
+            elementwise_mul_grad, ops::ElementWiseMulOpGrad);
+REGISTER_OP_CPU_KERNEL(
+    elementwise_mul,
+    ops::ElementWiseMulKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    elementwise_mul_grad,
+    ops::ElementWiseMulGradKernel<paddle::platform::CPUPlace, float>);
diff --git a/paddle/operators/elementwise_mul_op.cu b/paddle/operators/elementwise_mul_op.cu
new file mode 100644
index 0000000000..56f2087c22
--- /dev/null
+++ b/paddle/operators/elementwise_mul_op.cu
@@ -0,0 +1,25 @@
+/* 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. */
+
+#define EIGEN_USE_GPU
+#include "paddle/operators/elementwise_mul_op.h"
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_GPU_KERNEL(
+    elementwise_mul,
+    ops::ElementWiseMulKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    elementwise_mul_grad,
+    ops::ElementWiseMulGradKernel<paddle::platform::GPUPlace, float>);
diff --git a/paddle/operators/elementwise_mul_op.h b/paddle/operators/elementwise_mul_op.h
new file mode 100644
index 0000000000..6d58da580b
--- /dev/null
+++ b/paddle/operators/elementwise_mul_op.h
@@ -0,0 +1,183 @@
+/* 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 "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+/*
+ * Out = X ⊙ Y
+ * 1. shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1
+ *    pre=2, n=3*4, post=5
+ * 2. shape(X) = (2, 3, 4, 5), shape(Y) = (4,5)
+ *    pre=2*3, n=4*5, post=1
+ */
+
+inline void get_mid_dims(const framework::DDim& x_dims,
+                         const framework::DDim& y_dims, const int axis,
+                         int& pre, int& n, int& post) {
+  pre = 1;
+  n = 1;
+  post = 1;
+  for (int i = 0; i < axis; ++i) {
+    pre *= x_dims[i];
+  }
+
+  for (int i = 0; i < y_dims.size(); ++i) {
+    PADDLE_ENFORCE_EQ(x_dims[i + axis], y_dims[i],
+                      "Broadcast dimension mismatch.");
+    n *= y_dims[i];
+  }
+
+  for (int i = axis + y_dims.size(); i < x_dims.size(); ++i) {
+    post *= x_dims[i];
+  }
+}
+
+template <typename Place, typename T>
+class ElementWiseMulKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    using Tensor = framework::Tensor;
+
+    auto* x = ctx.Input<Tensor>("X");
+    auto* y = ctx.Input<Tensor>("Y");
+    auto* z = ctx.Output<Tensor>("Out");
+    z->mutable_data<T>(ctx.GetPlace());
+
+    auto x_e = framework::EigenVector<T>::Flatten(*x);
+    auto y_e = framework::EigenVector<T>::Flatten(*y);
+    auto z_e = framework::EigenVector<T>::Flatten(*z);
+
+    auto x_dims = x->dims();
+    auto y_dims = y->dims();
+    PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
+                      "Rank of first input must >= rank of second input.")
+
+    if (x_dims == y_dims || product(y_dims) == 1) {
+      z_e.device(ctx.GetEigenDevice<Place>()) = x_e * y_e;
+      return;
+    }
+
+    int axis = ctx.Attr<int>("axis");
+    axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
+    PADDLE_ENFORCE(axis >= 0 && axis < x_dims.size(),
+                   "Axis should be in range [0, x_dims)");
+
+    int pre, n, post;
+    get_mid_dims(x_dims, y_dims, axis, pre, n, post);
+    if (post == 1) {
+      auto y_bcast = y_e.reshape(Eigen::DSizes<int, 2>(1, n))
+                         .broadcast(Eigen::DSizes<int, 2>(pre, 1))
+                         .reshape(Eigen::DSizes<int, 1>(x_e.size()));
+      z_e.device(ctx.GetEigenDevice<Place>()) = x_e * y_bcast;
+      return;
+    } else {
+      auto y_bcast = y_e.reshape(Eigen::DSizes<int, 3>(1, n, 1))
+                         .broadcast(Eigen::DSizes<int, 3>(pre, 1, post))
+                         .reshape(Eigen::DSizes<int, 1>(x_e.size()));
+      z_e.device(ctx.GetEigenDevice<Place>()) = x_e * y_bcast;
+      return;
+    }
+  }
+};
+
+template <typename Place, typename T>
+class ElementWiseMulGradKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    using Tensor = framework::Tensor;
+
+    auto* x = ctx.Input<Tensor>("X");
+    auto* y = ctx.Input<Tensor>("Y");
+    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+
+    auto x_e = framework::EigenVector<T>::Flatten(*x);
+    auto y_e = framework::EigenVector<T>::Flatten(*y);
+    auto dout_e = framework::EigenVector<T>::Flatten(*dout);
+
+    auto x_dims = x->dims();
+    auto y_dims = y->dims();
+
+    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    if (dx) {
+      dx->mutable_data<T>(ctx.GetPlace());
+    }
+    if (dy) {
+      dy->mutable_data<T>(ctx.GetPlace());
+    }
+
+    if (x_dims == y_dims || product(y_dims) == 1) {
+      if (dx) {
+        auto dx_e = framework::EigenVector<T>::Flatten(*dx);
+        dx_e.device(ctx.GetEigenDevice<Place>()) = dout_e * y_e;
+      }
+
+      if (dy) {
+        auto dy_e = framework::EigenVector<T>::Flatten(*dy);
+        dy_e.device(ctx.GetEigenDevice<Place>()) = x_e * dout_e;
+      }
+      return;
+    }
+
+    int axis = ctx.Attr<int>("axis");
+    axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
+
+    int pre, n, post;
+    get_mid_dims(x_dims, y_dims, axis, pre, n, post);
+
+    // TODO(gongweibao): wrap reshape to a function.
+    if (post == 1) {
+      auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 2>(1, n))
+                           .broadcast(Eigen::DSizes<int, 2>(pre, 1))
+                           .reshape(Eigen::DSizes<int, 1>(x_e.size()));
+      if (dx) {
+        auto dx_e = framework::EigenVector<T>::Flatten(*dx);
+        dx_e.device(ctx.GetEigenDevice<Place>()) = dout_e * y_e_bcast;
+      }
+
+      if (dy) {
+        auto dy_e = framework::EigenVector<T>::Flatten(*dy);
+        dy_e.device(ctx.GetEigenDevice<Place>()) =
+            (x_e * dout_e)
+                .reshape(Eigen::DSizes<int, 2>(pre, n))
+                .sum(Eigen::array<int, 1>{{0}});
+      }
+      return;
+    } else {
+      auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 3>(1, n, 1))
+                           .broadcast(Eigen::DSizes<int, 3>(pre, 1, post))
+                           .reshape(Eigen::DSizes<int, 1>(x_e.size()));
+      if (dx) {
+        auto dx_e = framework::EigenVector<T>::Flatten(*dx);
+        dx_e.device(ctx.GetEigenDevice<Place>()) = dout_e * y_e_bcast;
+      }
+
+      if (dy) {
+        auto dy_e = framework::EigenVector<T>::Flatten(*dy);
+        dy_e.device(ctx.GetEigenDevice<Place>()) =
+            (x_e * dout_e)
+                .reshape(Eigen::DSizes<int, 3>(pre, n, post))
+                .sum(Eigen::array<int, 2>{{0, 2}});
+      }
+      return;
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/operators/fc_op.cc b/paddle/operators/fc_op.cc
new file mode 100644
index 0000000000..e5d0f3c372
--- /dev/null
+++ b/paddle/operators/fc_op.cc
@@ -0,0 +1,197 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/net_op.h"
+
+namespace paddle {
+namespace operators {
+
+class FCOp : public NetOp {
+ public:
+  FCOp(const std::string &type, const framework::VariableNameMap &inputs,
+       const framework::VariableNameMap &outputs,
+       const framework::AttributeMap &attrs)
+      : NetOp(type, inputs, outputs, attrs) {
+    PADDLE_ENFORCE(!Inputs("X").empty(),
+                   "Inputs(X) of FCOp should not be null.");
+    PADDLE_ENFORCE(!Inputs("W").empty(),
+                   "Inputs(W) of FCOp should not be null.");
+    PADDLE_ENFORCE(!Outputs("MulOut").empty(),
+                   "Outputs(MulOut) of FCOp should not be null.");
+    PADDLE_ENFORCE_NE(Output("Out"), framework::kEmptyVarName,
+                      "Output(Out) of FCOp should not be null.");
+
+    auto x = Inputs("X");
+    auto w = Inputs("W");
+    auto mul_out = Outputs("MulOut");
+    PADDLE_ENFORCE_EQ(
+        x.size(), w.size(),
+        "The size of inputs X(%d) should be the same as that of weights W(%d).",
+        x.size(), w.size());
+    PADDLE_ENFORCE_EQ(mul_out.size(), x.size(),
+                      "The size of intermediate mul_out(%d) should be the same "
+                      "as that of inputs X(%d).",
+                      mul_out.size(), x.size());
+
+    size_t n = x.size();
+    PADDLE_ENFORCE_GE(n, static_cast<size_t>(1),
+                      "The size of inputs X(%d) should be no less than 1.", n);
+
+    auto x_num_col_dims = Attr<std::vector<int>>("xNumColDims");
+
+    // Set all values or set no values (use the default value)
+    if (!x_num_col_dims.empty()) {
+      PADDLE_ENFORCE_EQ(x_num_col_dims.size(), n,
+                        "The size of attribute xNumColDims(%d) should be the "
+                        "same as that of inputs X(%d).",
+                        x_num_col_dims.size(), n);
+    } else {
+      x_num_col_dims.resize(n);
+      for (size_t i = 0; i < n; i++) {
+        x_num_col_dims[i] = 1;
+      }
+    }
+
+    // mul_out[i] = X[i] * W[i]
+    for (size_t i = 0; i < n; i++) {
+      framework::AttributeMap mul_attr;
+      mul_attr["x_num_col_dims"] = static_cast<int>(x_num_col_dims[i]);
+      mul_attr["y_num_col_dims"] = static_cast<int>(1);
+      AppendOp(
+          framework::OpRegistry::CreateOp("mul", {{"X", {x[i]}}, {"Y", {w[i]}}},
+                                          {{"Out", {mul_out[i]}}}, mul_attr));
+    }
+
+    // sum_out = X[0] * W[0] + ... + X[n-1] * W[n-1]
+    auto sum_out = mul_out[0];
+    if (n > 1) {
+      PADDLE_ENFORCE_NE(Output("SumOut"), framework::kEmptyVarName,
+                        "Output(SumOut) of FCOp should not be null when the "
+                        "size of Inputs(X) > 1.");
+
+      sum_out = Output("SumOut");
+      AppendOp(framework::OpRegistry::CreateOp("sum", {{"X", {mul_out}}},
+                                               {{"Out", {sum_out}}}, {}));
+    } else {
+      if (Output("SumOut") != framework::kEmptyVarName) {
+        this->Rename(Output("SumOut"), framework::kEmptyVarName);
+      }
+    }
+
+    // add_out = sum_out + b
+    auto b = Input("B");
+    auto add_out = sum_out;
+    if (b != framework::kEmptyVarName) {
+      PADDLE_ENFORCE_NE(
+          Output("AddOut"), framework::kEmptyVarName,
+          "Output(AddOut) of FCOp should not be null when Input(B) is set.");
+
+      add_out = Output("AddOut");
+      AppendOp(framework::OpRegistry::CreateOp(
+          "rowwise_add", {{"X", {sum_out}}, {"b", {Input("B")}}},
+          {{"Out", {add_out}}}, {}));
+    } else {
+      if (Output("AddOut") != framework::kEmptyVarName) {
+        this->Rename(Output("AddOut"), framework::kEmptyVarName);
+      }
+    }
+
+    auto activation = Attr<std::string>("activation");
+    AppendOp(framework::OpRegistry::CreateOp(activation, {{"X", {add_out}}},
+                                             {{"Y", {Output("Out")}}}, {}));
+    CompleteAddOp(false);
+  }
+};
+
+class FCOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  FCOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X",
+             "(A vector of Tensors) each input Tensor can be of arbitrary "
+             "dimension, and will be reshaped to a 2-D matrix of size "
+             "(minibatch, number_of_input_features) according to attribute "
+             "xNumColDims.")
+        .AsDuplicable();
+    AddInput("W",
+             "(A vector of Tensors) the weights of FC operator, a "
+             "vector of 2-D matrix of size "
+             "(number_of_input_features, number_of_neurons).")
+        .AsDuplicable();
+    AddInput("B",
+             "(Tensor) the bias of FC operator, a 1-D vector of size "
+             "number_of_neurons.");
+
+    AddOutput("Out",
+              "(Tensor) the activated output matrix of FC operator, a 2-D "
+              "matrix of size (minibatch, number_of_neurons).");
+    AddOutput("MulOut",
+              "(A vector of Tensors) the intermediate outputs of FC operator, "
+              "each Tensor saving the product of X_i * W_i.")
+        .AsIntermediate()
+        .AsDuplicable();
+    AddOutput(
+        "SumOut",
+        "(Tensor) the intermediate output of FC operator, "
+        "saving the sum of the products of X and W, that is sum{X_i * W_i}.")
+        .AsIntermediate();
+    AddOutput("AddOut",
+              "(Tensor) the non-actived output of FC operator, "
+              "saving sum{X_i * W_i} + B.")
+        .AsIntermediate();
+    AddAttr<std::string>(
+        "activation",
+        "(string, default identity) the activation type of FC operator.")
+        .SetDefault("identity")
+        .InEnum({"identity", "sigmoid", "softmax"});
+    AddAttr<std::vector<int>>(
+        "xNumColDims",
+        "(std::vector<int>) The inputs Tensors of FC operator can be of "
+        "more than 2 dimensions. In that case, each input Tensor `X_i` will be "
+        "reshaped to a 2-D matrix. The matrix's first dimension "
+        "(the length of column) will be the product of `X_i`'s last "
+        "`xNumColDims_i` dimensions, that is "
+        "`X_i.dims[0] x ... x X_i.dims[xNumColDims_i - 1]`. "
+        "The matrix's second dimension (the length of row) will be the product "
+        "of `X_i`'s first `rank - xNumColDims_i` dimensions, that is "
+        "`X_i.dims[xNumColDims_i] x ... x X_i.dims[rank - 1]`)")
+        .SetDefault(std::vector<int>{});
+
+    AddComment(R"DOC(
+Fully Connected Operator, known as Fully Connected Layer or Inner Product Layer
+in Convolutional Neural Networks. Neurons in a fully connected layer have
+full connections to all activations in the previous layer.
+It computes an inner product of a set of
+learned weights with a matrix multiplication followed by a bias offset
+(optionally).
+
+Equation:
+  Out = Act(sum_n{X_i * W_i} + B)
+
+where X_i is Tensor that will be reshaped to a 2-D matrix of size (M x K),
+usually M is the minibatch size and K is the number of input features.
+W_i is a 2-D matrix of size (K x N), where N means the number of neurons
+in the fully connected layer. B is a 1-D vector of size N.
+Thus, the output Out is a 2-D matrix of size (M x N).
+Activation type can be set to `identity` (default), `sigmoid` or `softmax`.
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(fc, ops::FCOp, ops::FCOpMaker);
diff --git a/paddle/operators/fill_zeros_like_op.cc b/paddle/operators/fill_zeros_like_op.cc
index 9d51f6e3a1..ba7857cc65 100644
--- a/paddle/operators/fill_zeros_like_op.cc
+++ b/paddle/operators/fill_zeros_like_op.cc
@@ -23,7 +23,14 @@ class FillZerosLikeOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    ctx.Output<framework::Tensor>("Dst")->Resize(
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.InputVar("Src"),
+        "Input(Src) of FillZerosLikeOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Dst"),
+        "Output(Dst) of FillZerosLikeOp should not be null.");
+
+    ctx.Output<framework::LoDTensor>("Dst")->Resize(
         ctx.Input<framework::Tensor>("Src")->dims());
   }
 };
diff --git a/paddle/operators/gather_op.cc b/paddle/operators/gather_op.cc
index 123bed296c..d445b61c16 100644
--- a/paddle/operators/gather_op.cc
+++ b/paddle/operators/gather_op.cc
@@ -24,11 +24,18 @@ class GatherOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of GatherOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Index"),
+                            "Input(Index) of GatherOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of GatherOp should not be null.");
+
     int batch_size = ctx.Input<Tensor>("Index")->dims()[0];
     PADDLE_ENFORCE_GE(batch_size, 0, "Batch size must be >0");
     framework::DDim output_dims(ctx.Input<Tensor>("X")->dims());
     output_dims[0] = batch_size;
-    ctx.Output<Tensor>("Out")->Resize(output_dims);
+    ctx.Output<framework::LoDTensor>("Out")->Resize(output_dims);
   }
 };
 
@@ -38,7 +45,7 @@ class GatherGradOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    auto X_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto X_grad = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
     auto X = ctx.Input<Tensor>("X");
 
     X_grad->Resize(X->dims());
diff --git a/paddle/operators/gaussian_random_op.cc b/paddle/operators/gaussian_random_op.cc
index 6574880c0e..c0e161bbc0 100644
--- a/paddle/operators/gaussian_random_op.cc
+++ b/paddle/operators/gaussian_random_op.cc
@@ -31,7 +31,7 @@ class CPUGaussianRandomKernel : public framework::OpKernel {
     }
     engine.seed(seed);
     std::normal_distribution<T> dist(mean, std);
-    int64_t size = framework::product(tensor->dims());
+    int64_t size = tensor->numel();
     for (int64_t i = 0; i < size; ++i) {
       data[i] = dist(engine);
     }
@@ -43,8 +43,12 @@ class GaussianRandomOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext& context) const override {
-    auto* tensor = context.Output<framework::Tensor>("Out");
+  void InferShape(const framework::InferShapeContext& ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Out"),
+        "Output(Out) of GaussianRandomOp should not be null.");
+
+    auto* tensor = ctx.Output<framework::LoDTensor>("Out");
     auto dims = Attr<std::vector<int>>("dims");
     std::vector<int64_t> temp;
     temp.reserve(dims.size());
diff --git a/paddle/operators/gaussian_random_op.cu b/paddle/operators/gaussian_random_op.cu
index d9dbc1dcfe..2d63b30499 100644
--- a/paddle/operators/gaussian_random_op.cu
+++ b/paddle/operators/gaussian_random_op.cu
@@ -50,8 +50,8 @@ class GPUGaussianRandomKernel : public framework::OpKernel {
     T mean = static_cast<T>(context.Attr<float>("mean"));
     T std = static_cast<T>(context.Attr<float>("std"));
     thrust::counting_iterator<unsigned int> index_sequence_begin(0);
-    ssize_t N = framework::product(tensor->dims());
-    thrust::transform(index_sequence_begin, index_sequence_begin + N,
+    int64_t size = tensor->numel();
+    thrust::transform(index_sequence_begin, index_sequence_begin + size,
                       thrust::device_ptr<T>(data),
                       GaussianGenerator<T>(mean, std, seed));
   }
diff --git a/paddle/operators/identity_op.cc b/paddle/operators/identity_op.cc
index 7d9d4fa519..2cc632205e 100644
--- a/paddle/operators/identity_op.cc
+++ b/paddle/operators/identity_op.cc
@@ -27,7 +27,7 @@ class IdentityOpMaker : public framework::OpProtoAndCheckerMaker {
                   framework::OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "The input tensor of identity operator.");
-    AddOutput("Out", "The output tensor of identity operator.");
+    AddOutput("Y", "The output tensor of identity operator.");
     AddComment(R"DOC(
 The identity operator is an alias of the scale operator
 with the attribute scale fixed to 1.0.
@@ -42,9 +42,15 @@ class IdentityOp : public NetOp {
              const framework::VariableNameMap &outputs,
              const framework::AttributeMap &attrs)
       : NetOp(type, inputs, outputs, attrs) {
+    PADDLE_ENFORCE_NE(Input("X"), framework::kEmptyVarName,
+                      "Input(X) of IdentityOp should not be null.");
+    PADDLE_ENFORCE_NE(Output("Y"), framework::kEmptyVarName,
+                      "Output(Y) of IdentityOp should not be null.");
+
     AppendOp(framework::OpRegistry::CreateOp(
-        "scale", {{"X", {Input("X")}}}, {{"Out", {Output("Out")}}},
+        "scale", {{"X", {Input("X")}}}, {{"Out", {Output("Y")}}},
         {{"scale", static_cast<AttrType>(1)}}));
+    CompleteAddOp(false);
   }
 };
 
diff --git a/paddle/operators/lookup_table_op.cc b/paddle/operators/lookup_table_op.cc
index 94d40890a7..07f6dfabca 100644
--- a/paddle/operators/lookup_table_op.cc
+++ b/paddle/operators/lookup_table_op.cc
@@ -22,10 +22,17 @@ class LookupTableOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &context) const override {
-    auto table_t = context.Input<Tensor>("W");
-    auto ids_t = context.Input<Tensor>("Ids");
-    auto output_t = context.Output<Tensor>("Out");
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("W"),
+                            "Input(W) of LookupTableOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Ids"),
+                            "Input(Ids) of LookupTableOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of LookupTableOp should not be null.");
+
+    auto table_t = ctx.Input<Tensor>("W");
+    auto ids_t = ctx.Input<Tensor>("Ids");
+    auto output_t = ctx.Output<framework::LoDTensor>("Out");
 
     output_t->Resize({ids_t->dims()[0], table_t->dims()[1]});
   }
@@ -56,7 +63,8 @@ class LookupTableOpGrad : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext &context) const override {
     auto table = context.Input<Tensor>("W");
-    auto d_table = context.Output<Tensor>(framework::GradVarName("W"));
+    auto d_table =
+        context.Output<framework::LoDTensor>(framework::GradVarName("W"));
     d_table->Resize(table->dims());
   }
 };
diff --git a/paddle/operators/lookup_table_op.cu b/paddle/operators/lookup_table_op.cu
index 27eee3436a..7083440467 100644
--- a/paddle/operators/lookup_table_op.cu
+++ b/paddle/operators/lookup_table_op.cu
@@ -70,7 +70,7 @@ class LookupTableCUDAKernel : public framework::OpKernel {
 
     size_t N = table_t->dims()[0];
     size_t D = table_t->dims()[1];
-    size_t K = product(ids_t->dims());
+    size_t K = ids_t->numel();
     auto ids = ids_t->data<int32_t>();
     auto table = table_t->data<T>();
     auto output = output_t->mutable_data<T>(context.GetPlace());
@@ -91,7 +91,7 @@ class LookupTableGradCUDAKernel : public framework::OpKernel {
 
     int N = d_table_t->dims()[0];
     int D = d_table_t->dims()[1];
-    int K = product(ids_t->dims());
+    int K = ids_t->numel();
     const int32_t* ids = ids_t->data<int32_t>();
     const T* d_output = d_output_t->data<T>();
     T* d_table = d_table_t->mutable_data<T>(context.GetPlace());
diff --git a/paddle/operators/lookup_table_op.h b/paddle/operators/lookup_table_op.h
index 877b36cef4..a1298906dd 100644
--- a/paddle/operators/lookup_table_op.h
+++ b/paddle/operators/lookup_table_op.h
@@ -35,7 +35,7 @@ class LookupTableKernel : public framework::OpKernel {
     auto ids = ids_t->data<int32_t>();
     auto table = table_t->data<T>();
     auto output = output_t->mutable_data<T>(context.GetPlace());
-    for (ssize_t i = 0; i < product(ids_t->dims()); ++i) {
+    for (int64_t i = 0; i < ids_t->numel(); ++i) {
       PADDLE_ENFORCE_LT(ids[i], N);
       PADDLE_ENFORCE_GE(ids[i], 0);
       memcpy(output + i * D, table + ids[i] * D, D * sizeof(T));
@@ -61,7 +61,7 @@ class LookupTableGradKernel : public framework::OpKernel {
     t.device(context.GetEigenDevice<platform::CPUPlace>()) =
         t.constant(static_cast<T>(0));
 
-    for (ssize_t i = 0; i < product(ids_t->dims()); ++i) {
+    for (int64_t i = 0; i < ids_t->numel(); ++i) {
       PADDLE_ENFORCE_LT(ids[i], N);
       PADDLE_ENFORCE_GE(ids[i], 0);
       for (int j = 0; j < D; ++j) {
diff --git a/paddle/operators/math/im2col_test.cc b/paddle/operators/math/im2col_test.cc
index 186a33edce..4f380388b1 100644
--- a/paddle/operators/math/im2col_test.cc
+++ b/paddle/operators/math/im2col_test.cc
@@ -119,4 +119,4 @@ TEST(math, im2col) {
 #ifndef PADDLE_ONLY_CPU
   testIm2col<paddle::platform::GPUPlace>();
 #endif
-}
\ No newline at end of file
+}
diff --git a/paddle/operators/mean_op.cc b/paddle/operators/mean_op.cc
index d3d0e55a67..7d7eeb59a2 100644
--- a/paddle/operators/mean_op.cc
+++ b/paddle/operators/mean_op.cc
@@ -24,8 +24,10 @@ class MeanOp : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
-                            "Input of MeanOp must be initialized.");
-    ctx.Output<Tensor>("Out")->Resize({1});
+                            "Input(X) of MeanOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of MeanOp should not be null.");
+    ctx.Output<framework::LoDTensor>("Out")->Resize({1});
   }
 };
 
@@ -45,7 +47,7 @@ class MeanGradOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    ctx.Output<Tensor>(framework::GradVarName("X"))
+    ctx.Output<framework::LoDTensor>(framework::GradVarName("X"))
         ->Resize(ctx.Input<Tensor>("X")->dims());
   }
 };
diff --git a/paddle/operators/mean_op.h b/paddle/operators/mean_op.h
index 9848af280b..ce31e178d8 100644
--- a/paddle/operators/mean_op.h
+++ b/paddle/operators/mean_op.h
@@ -49,12 +49,11 @@ class MeanGradKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
     auto OG = context.Input<Tensor>(framework::GradVarName("Out"));
-    PADDLE_ENFORCE(framework::product(OG->dims()) == 1,
-                   "Mean Gradient should be scalar");
+    PADDLE_ENFORCE(OG->numel() == 1, "Mean Gradient should be scalar");
     auto IG = context.Output<Tensor>(framework::GradVarName("X"));
     IG->mutable_data<T>(context.GetPlace());
 
-    T ig_size = (T)framework::product(IG->dims());
+    T ig_size = static_cast<T>(IG->numel());
     Eigen::DSizes<int, 1> bcast(ig_size);
 
     EigenVector<T>::Flatten(*IG).device(context.GetEigenDevice<Place>()) =
diff --git a/paddle/operators/minus_op.cc b/paddle/operators/minus_op.cc
index 069fb5e1ab..a97bbecdca 100644
--- a/paddle/operators/minus_op.cc
+++ b/paddle/operators/minus_op.cc
@@ -27,14 +27,20 @@ class MinusOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of MinusOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
+                            "Input(Y) of MinusOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of MinusOp should not be null.");
+
     auto *left_tensor = ctx.Input<framework::Tensor>("X");
     auto *right_tensor = ctx.Input<framework::Tensor>("Y");
 
     PADDLE_ENFORCE_EQ(
-        framework::product(left_tensor->dims()),
-        framework::product(right_tensor->dims()),
+        left_tensor->numel(), right_tensor->numel(),
         "Minus operator must take two tensor with same num of elements");
-    ctx.Output<framework::Tensor>("Out")->Resize(left_tensor->dims());
+    ctx.Output<framework::LoDTensor>("Out")->Resize(left_tensor->dims());
   }
 };
 
@@ -65,7 +71,7 @@ class MinusGradOp : public NetOp {
 
     // x_grad = out_grad
     AppendOp(framework::OpRegistry::CreateOp("identity", {{"X", {out_grad}}},
-                                             {{"Out", {x_grad}}}, {}));
+                                             {{"Y", {x_grad}}}, {}));
 
     framework::AttributeMap scale_attr;
     scale_attr["scale"] = static_cast<AttrType>(-1);
@@ -78,8 +84,6 @@ class MinusGradOp : public NetOp {
 }  // namespace operators
 }  // namespace paddle
 
-USE_OP(scale);
-USE_NO_KERNEL_OP(identity);
 namespace ops = paddle::operators;
 REGISTER_OP(minus, ops::MinusOp, ops::MinusOpMaker, minus_grad,
             ops::MinusGradOp<float>);
diff --git a/paddle/operators/mul_op.cc b/paddle/operators/mul_op.cc
index 710a56a0e8..b6d320b415 100644
--- a/paddle/operators/mul_op.cc
+++ b/paddle/operators/mul_op.cc
@@ -18,6 +18,7 @@ namespace paddle {
 namespace operators {
 
 using framework::Tensor;
+using framework::LoDTensor;
 
 class MulOp : public framework::OperatorWithKernel {
  public:
@@ -25,6 +26,13 @@ class MulOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of MulOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
+                            "Input(Y) of MulOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of MulOp should not be null.");
+
     auto x_dims = ctx.Input<Tensor>("X")->dims();
     auto y_dims = ctx.Input<Tensor>("Y")->dims();
     int x_num_col_dims = Attr<int>("x_num_col_dims");
@@ -45,7 +53,8 @@ class MulOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(
         x_mat_dims[1], y_mat_dims[0],
         "First matrix's width must be equal with second matrix's height.");
-    ctx.Output<Tensor>("Out")->Resize({x_mat_dims[0], y_mat_dims[1]});
+    ctx.Output<framework::LoDTensor>("Out")->Resize(
+        {x_mat_dims[0], y_mat_dims[1]});
   }
 };
 
@@ -94,8 +103,10 @@ class MulOpGrad : public framework::OperatorWithKernel {
     auto x_dims = ctx.Input<Tensor>("X")->dims();
     auto y_dims = ctx.Input<Tensor>("Y")->dims();
     auto out_dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
-    auto *x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
-    auto *y_grad = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    auto *x_grad =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto *y_grad =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("Y"));
 
     auto x_mat_dims =
         framework::flatten_to_2d(x_dims, Attr<int>("x_num_col_dims"));
diff --git a/paddle/operators/name_convention.md b/paddle/operators/name_convention.md
new file mode 100644
index 0000000000..379385dc5d
--- /dev/null
+++ b/paddle/operators/name_convention.md
@@ -0,0 +1,61 @@
+## Operator's Parameter Name Convention
+
+To make the operator document itself more clear, we recommend operator names obey the listing conventions.
+
+### OpProtoMaker names
+
+When defining an operator in Paddle, a corresponding [OpProtoMaker](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h#L170) (TODO: OpProtoMaker Doc)need to be defined. All the Input/Output and Attributes will write into the [OpProto](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto#L61) , and will be used in client language to create operator. 
+
+- Input/Output.
+  - Input/Output names follow the **CamelCase**. e.g. `X`, `Y`, `Matrix`, `LastAxisInMatrix`. Input/Output much more like Variables, we prefer to meaningful English words. 
+  - If an operator's Input/Output are tensors in math, not match to any meaningful words, input name should starts from `X`. e.g. `X`, `Y`, and output name should starts from `Out`. e.g. `Out`. This rule intends making operators which have few inputs/outputs unified.
+
+- Attribute.
+  - Attribute name follows the **camelCase**. e.g. `x`, `y`, `axis`, `rowwiseMatrix`. Also, attribute name prefers to meaningful English words.
+
+- Comments.
+  - Input/Output/Attr comment follow the format of **(type,default value) usage**, corresponding to which type it can be and how it will be used in the operator. e.g.  Attribute in Accumulator`"gamma" `,`(float, default 1.0) Accumulation multiplier`.
+  - Operator comment format of` R"DOC(your comment here)DOC"`. You should explain the input/output of the operator first. If there is math calculation in this operator, you should write the equation in the comment. e.g. `Out = X + Y`. 
+
+- Order.
+  - Follow the order of Input/Output, then Attribute, then Comments. See the example in best practice.
+
+### Best Practice
+
+Here we give some examples to show how these rules will be used.
+
+- The operator has one input, one output. e.g.`relu`, inputs: `X`, outputs: `Out`. 
+
+- The operator has two input, one output. e.g. `rowwise_add`, inputs : `X`, `Y`, outputs : `Out`.
+
+- The operator contains attribute. e.g. `cosine`, inputs : `X`, `axis`, outputs : `Out`.
+
+  We give a full example of Accumulator Operator.
+
+```c++
+class AccumulateOpMaker : public framework::OpProtoAndCheckerMaker {
+public:
+  AccumulateOpMaker(framework::OpProto *proto,
+                            framework::OpAttrChecker *op_checker)
+    : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "(Tensor) The input tensor that has to be accumulated to the output tensor. 
+    If the output size is not the same as input size, 
+    the output tensor is first reshaped and initialized to zero, and only then, accumulation is done.");
+    AddOutput("Out", "(Tensor) Accumulated output tensor");
+    AddAttr<float>("gamma", "(float, default 1.0) Accumulation multiplier").SetDefault(1.0f);
+    AddComment(R"DOC(
+Accumulate operator accumulates the input tensor to the output tensor. If the
+output tensor already has the right size, we add to it; otherwise, we first
+initialize the output tensor to all zeros, and then do accumulation. Any
+further calls to the operator, given that no one else fiddles with the output
+in the interim, will do simple accumulations.
+Accumulation is done as shown:
+
+Out = 1*X + gamma*Out
+
+where X is the input tensor, Out is the output tensor and gamma is the multiplier
+argument.
+)DOC");
+  }
+};
+```
diff --git a/paddle/operators/cross_entropy_op.cc b/paddle/operators/onehot_cross_entropy_op.cc
similarity index 81%
rename from paddle/operators/cross_entropy_op.cc
rename to paddle/operators/onehot_cross_entropy_op.cc
index ab1e1c101a..f38be3549f 100644
--- a/paddle/operators/cross_entropy_op.cc
+++ b/paddle/operators/onehot_cross_entropy_op.cc
@@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/operators/cross_entropy_op.h"
+#include "paddle/operators/onehot_cross_entropy_op.h"
 
 namespace paddle {
 namespace operators {
@@ -23,13 +23,23 @@ class OnehotCrossEntropyOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.InputVar("X"),
+        "Input(X) of OnehotCrossEntropyOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.InputVar("label"),
+        "Input(label) of OnehotCrossEntropyOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Y"),
+        "Output(Y) of OnehotCrossEntropyOp should not be null.");
+
     auto *X = ctx.Input<Tensor>("X");
     auto *label = ctx.Input<Tensor>("label");
 
     PADDLE_ENFORCE_EQ(X->dims().size(), 2, "X's dimension must be 2.");
     PADDLE_ENFORCE_EQ(label->dims().size(), 1, "label's dimension must be 1.");
     PADDLE_ENFORCE_EQ(X->dims()[0], label->dims()[0]);
-    ctx.Output<Tensor>("Y")->Resize({X->dims()[0]});
+    ctx.Output<framework::LoDTensor>("Y")->Resize({X->dims()[0], 1});
   }
 };
 
@@ -39,7 +49,7 @@ class OnehotCrossEntropyGradientOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    auto dX = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto dX = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
     auto X = ctx.Input<Tensor>("X");
 
     dX->Resize(X->dims());
diff --git a/paddle/operators/cross_entropy_op.cu b/paddle/operators/onehot_cross_entropy_op.cu
similarity index 100%
rename from paddle/operators/cross_entropy_op.cu
rename to paddle/operators/onehot_cross_entropy_op.cu
diff --git a/paddle/operators/cross_entropy_op.h b/paddle/operators/onehot_cross_entropy_op.h
similarity index 100%
rename from paddle/operators/cross_entropy_op.h
rename to paddle/operators/onehot_cross_entropy_op.h
diff --git a/paddle/operators/pad_op.cc b/paddle/operators/pad_op.cc
new file mode 100644
index 0000000000..a0b1c6b631
--- /dev/null
+++ b/paddle/operators/pad_op.cc
@@ -0,0 +1,118 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/operators/pad_op.h"
+
+namespace paddle {
+namespace operators {
+
+using framework::Tensor;
+
+class PadOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of PadOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of PadOp should not be null.");
+
+    auto x_dim = ctx.Input<Tensor>("X")->dims();
+    auto paddings = Attr<std::vector<int>>("paddings");
+    PADDLE_ENFORCE_EQ(x_dim.size() * 2, int64_t(paddings.size()),
+                      "Size of paddings should be equal to 2 * dimension size "
+                      "of input tensor.");
+    std::vector<int64_t> out_dims(x_dim.size());
+    for (int i = 0; i < x_dim.size(); ++i) {
+      out_dims[i] = x_dim[i] + paddings[i * 2] + paddings[i * 2 + 1];
+    }
+    ctx.Output<framework::LoDTensor>("Out")->Resize(
+        framework::make_ddim(out_dims));
+  }
+};
+
+class PadOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  PadOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X",
+             "The input of pad op. "
+             "The input should be a k-D tensor(k > 0 and k < 7)");
+    AddOutput("Out",
+              "The output of pad op."
+              "A tensor with the same shape as X.")
+        .NotInGradient();
+    AddComment(R"DOC(
+Pad input into output, as specified by paddings and pad_value. The input should be a k-D tensor(k > 0 and k < 7). As an example:
+
+Given:
+
+X = [[1, 2],
+   [3, 4]]
+
+and 
+
+paddings = [0, 1, 1, 2]
+
+and
+ 
+pad_value = 0 
+
+then we get 
+
+Out = [[0, 1, 2, 0, 0]
+       [0, 3, 4, 0, 0]
+       [0, 0, 0, 0, 0]]
+)DOC");
+    AddAttr<std::vector<int>>(
+        "paddings",
+        "A list<int> to describes padding rules for each dimension."
+        " For 2-D image tensor, paddings=[0, 1, 2, 3] means"
+        " padding 0 row to top, 1 row to bottom, 2 columns to left"
+        " and 3 columns to right.Size of paddings should be equal to"
+        " 2 * dimension size of input tensor.");
+    AddAttr<float>("pad_value",
+                   "(float) default to 0; "
+                   "The value to fill padded areas.")
+        .SetDefault(0.0f);
+  }
+};
+
+class PadOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
+                            "Input(Out@GRAD) should not be null");
+    auto x_dims = ctx.Input<Tensor>("X")->dims();
+    auto *x_g = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    if (x_g != nullptr) {
+      x_g->Resize(x_dims);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(pad, ops::PadOp, ops::PadOpMaker, pad_grad, ops::PadOpGrad);
+REGISTER_OP_CPU_KERNEL(pad, ops::PadKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(pad_grad,
+                       ops::PadGradKernel<paddle::platform::CPUPlace, float>);
diff --git a/paddle/operators/pad_op.cu b/paddle/operators/pad_op.cu
new file mode 100644
index 0000000000..555a7dba23
--- /dev/null
+++ b/paddle/operators/pad_op.cu
@@ -0,0 +1,21 @@
+/* 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. */
+
+#define EIGEN_USE_GPU
+#include "paddle/operators/pad_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(pad, ops::PadKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(pad_grad,
+                       ops::PadGradKernel<paddle::platform::GPUPlace, float>);
diff --git a/paddle/operators/pad_op.h b/paddle/operators/pad_op.h
new file mode 100644
index 0000000000..2cc3b945ae
--- /dev/null
+++ b/paddle/operators/pad_op.h
@@ -0,0 +1,132 @@
+/* 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 "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename T, size_t D, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>;
+
+template <typename Place, typename T, size_t D>
+void PadFunction(const framework::ExecutionContext& context) {
+  auto pads = context.Attr<std::vector<int>>("paddings");
+  Eigen::array<std::pair<int, int>, D> paddings;
+  for (size_t i = 0; i < paddings.size(); ++i) {
+    paddings[i].first = pads[i * 2];
+    paddings[i].second = pads[i * 2 + 1];
+  }
+  T pad_value = context.Attr<T>("pad_value");
+
+  auto* x = context.Input<Tensor>("X");
+  auto* out = context.Output<Tensor>("Out");
+  out->mutable_data<T>(context.GetPlace());
+
+  auto x_tensor = EigenTensor<T, D>::From(*x);
+  auto out_tensor = EigenTensor<T, D>::From(*out);
+  auto place = context.GetEigenDevice<Place>();
+  out_tensor.device(place) = x_tensor.pad(paddings, pad_value);
+}
+
+template <typename Place, typename T>
+class PadKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    int rank = context.Input<Tensor>("X")->dims().size();
+    switch (rank) {
+      case 1:
+        PadFunction<Place, T, 1>(context);
+        break;
+      case 2:
+        PadFunction<Place, T, 2>(context);
+        break;
+      case 3:
+        PadFunction<Place, T, 3>(context);
+        break;
+      case 4:
+        PadFunction<Place, T, 4>(context);
+        break;
+      case 5:
+        PadFunction<Place, T, 5>(context);
+        break;
+      case 6:
+        PadFunction<Place, T, 6>(context);
+        break;
+      default:
+        PADDLE_THROW(
+            "PadOp only support tensors with no more than 6 dimensions.");
+    }
+  }
+};
+
+template <typename Place, typename T, size_t D>
+void PadGradFunction(const framework::ExecutionContext& context) {
+  auto pads = context.Attr<std::vector<int>>("paddings");
+  Eigen::array<std::pair<int, int>, D> paddings;
+  for (size_t i = 0; i < paddings.size(); ++i) {
+    paddings[i].first = -pads[i * 2];
+    paddings[i].second = -pads[i * 2 + 1];
+  }
+  auto* d_out = context.Input<Tensor>(framework::GradVarName("Out"));
+  auto* d_x = context.Output<Tensor>(framework::GradVarName("X"));
+  if (d_x != nullptr) {
+    d_x->mutable_data<T>(context.GetPlace());
+    auto d_x_tensor = EigenTensor<T, D>::From(*d_x);
+    auto d_out_tensor = EigenTensor<T, D>::From(*d_out);
+    auto place = context.GetEigenDevice<Place>();
+    d_x_tensor.device(place) = d_out_tensor.pad(paddings, 0);
+  }
+}
+
+template <typename Place, typename T>
+class PadGradKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    size_t rank =
+        context.Input<Tensor>(framework::GradVarName("Out"))->dims().size();
+    switch (rank) {
+      case 1:
+        PadGradFunction<Place, T, 1>(context);
+        break;
+      case 2:
+        PadGradFunction<Place, T, 2>(context);
+        break;
+      case 3:
+        PadGradFunction<Place, T, 3>(context);
+        break;
+      case 4:
+        PadGradFunction<Place, T, 4>(context);
+        break;
+      case 5:
+        PadGradFunction<Place, T, 5>(context);
+        break;
+      case 6:
+        PadGradFunction<Place, T, 6>(context);
+        break;
+      default:
+        PADDLE_THROW(
+            "PadOp only support tensors with no more than 6 dimensions.");
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/operators/recurrent_op.cc b/paddle/operators/recurrent_op.cc
index e826703c60..d3413d7cb9 100644
--- a/paddle/operators/recurrent_op.cc
+++ b/paddle/operators/recurrent_op.cc
@@ -26,10 +26,11 @@ namespace operators {
 using Scope = framework::Scope;
 using Variable = framework::Variable;
 using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
 
 void RecurrentAlgorithm::InferShape(const Scope& scope) const {
   seq_len_ = scope.FindVar((arg_->inlinks[0]).external)
-                 ->GetMutable<Tensor>()
+                 ->GetMutable<LoDTensor>()
                  ->dims()[0];
   CreateScopes(scope);
   auto step_scopes = GetStepScopes(scope);
@@ -88,7 +89,7 @@ void RecurrentAlgorithm::CreateScopes(const Scope& scope) const {
         // the weight are located in parent scope
         for (auto& var_name : input.second) {
           if (!step_scope.FindVar(var_name)) {
-            step_scope.NewVar(var_name)->GetMutable<Tensor>();
+            step_scope.NewVar(var_name)->GetMutable<LoDTensor>();
           }
         }
       }
@@ -106,11 +107,12 @@ void RecurrentAlgorithm::CreateScopes(const Scope& scope) const {
 void RecurrentAlgorithm::InitMemories(Scope* step_scope,
                                       bool infer_shape_mode) const {
   for (auto& attr : arg_->memories) {
-    Tensor* pre_mem = step_scope->NewVar(attr.pre_var)->GetMutable<Tensor>();
+    auto* pre_mem = step_scope->NewVar(attr.pre_var)->GetMutable<LoDTensor>();
     PADDLE_ENFORCE(step_scope->FindVar(attr.boot_var) != nullptr,
                    "memory [%s]'s boot variable [%s] not exists", attr.var,
                    attr.boot_var);
-    Tensor* boot_mem = step_scope->FindVar(attr.boot_var)->GetMutable<Tensor>();
+    auto* boot_mem =
+        step_scope->FindVar(attr.boot_var)->GetMutable<LoDTensor>();
     if (infer_shape_mode) {
       pre_mem->Resize(boot_mem->dims());
       PADDLE_ENFORCE_EQ(pre_mem->dims().size(), 2);
@@ -192,9 +194,9 @@ void RecurrentGradientAlgorithm::LinkBootMemoryGradients(
                    "memory variable [%s] does not exists", attr.var);
     PADDLE_ENFORCE(step_scope->FindVar(attr.boot_var) != nullptr,
                    "boot variable [%s] does not exists", attr.boot_var);
-    Tensor* mem_grad = step_scope->NewVar(attr.var)->GetMutable<Tensor>();
-    Tensor* boot_mem_grad =
-        step_scope->NewVar(attr.boot_var)->GetMutable<Tensor>();
+    auto* mem_grad = step_scope->NewVar(attr.var)->GetMutable<LoDTensor>();
+    auto* boot_mem_grad =
+        step_scope->NewVar(attr.boot_var)->GetMutable<LoDTensor>();
     if (infer_shape_mode) {
       boot_mem_grad->Resize(mem_grad->dims());
     } else {
@@ -205,7 +207,7 @@ void RecurrentGradientAlgorithm::LinkBootMemoryGradients(
 
 void RecurrentGradientAlgorithm::InferShape(const Scope& scope) const {
   seq_len_ = scope.FindVar((arg_->inlinks[0]).external)
-                 ->GetMutable<Tensor>()
+                 ->GetMutable<LoDTensor>()
                  ->dims()[0];
   auto step_scopes = GetStepScopes(scope);
   rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_,
diff --git a/paddle/operators/reshape_op.cc b/paddle/operators/reshape_op.cc
new file mode 100644
index 0000000000..0d05e34414
--- /dev/null
+++ b/paddle/operators/reshape_op.cc
@@ -0,0 +1,111 @@
+
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/operators/reshape_op.h"
+
+namespace paddle {
+namespace operators {
+
+class ReshapeOp : public framework::OperatorWithKernel {
+ public:
+  ReshapeOp(const std::string &type, const framework::VariableNameMap &inputs,
+            const framework::VariableNameMap &outputs,
+            const framework::AttributeMap &attrs)
+      : OperatorWithKernel(type, inputs, outputs, attrs) {}
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    // input check
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of ReshapeOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of ReshapeOp should not be null.");
+
+    auto shape = ctx.Attr<std::vector<int>>("shape");
+    PADDLE_ENFORCE(shape.size() > 0, "Attr(shape) shouldn't be empty.");
+    for (auto dim : shape) {
+      PADDLE_ENFORCE(dim > 0, "Each dimension of shape must be positive.");
+    }
+    // capacity check
+    int64_t capacity =
+        std::accumulate(shape.begin(), shape.end(), 1, std::multiplies<int>());
+    auto *in = ctx.Input<framework::Tensor>("X");
+    int64_t in_size = framework::product(in->dims());
+    PADDLE_ENFORCE_EQ(capacity, in_size,
+                      "The size of Input(X) mismatches with Attr(shape).");
+    // resize output
+    std::vector<int64_t> shape_int64(shape.size(), 0);
+    std::transform(shape.begin(), shape.end(), shape_int64.begin(),
+                   [](int a) { return static_cast<int64_t>(a); });
+    auto out_dims = framework::make_ddim(shape_int64);
+    ctx.Output<framework::LoDTensor>("Out")->Resize(out_dims);
+  }
+};
+
+class ReshapeOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  ReshapeOpMaker(framework::OpProto *proto,
+                 framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "The input tensor of reshape operator.");
+    AddOutput("Out", "The output tensor of reshape operator.");
+    AddAttr<std::vector<int>>("shape", "Target shape of reshape operator.");
+    AddComment(R"DOC(Reshape operator
+
+Reshape Input(X) into the shape specified by Attr(shape).
+
+An example:
+Given a 2-D tensor X with 2 rows and 2 columns
+
+    [[1, 2], [3, 4]]
+
+with target shape = [1, 4], the reshape operator will transform 
+the tensor X into a 1-D tensor:
+
+    [1, 2, 3, 4]
+
+)DOC");
+  }
+};
+
+class ReshapeGradOp : public framework::OperatorWithKernel {
+ public:
+  ReshapeGradOp(const std::string &type,
+                const framework::VariableNameMap &inputs,
+                const framework::VariableNameMap &outputs,
+                const framework::AttributeMap &attrs)
+      : OperatorWithKernel(type, inputs, outputs, attrs) {}
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) shouldn't be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
+                            "Input(Out@GRAD) shouldn't be null.");
+    auto dims = ctx.Input<framework::Tensor>("X")->dims();
+    auto *d_in = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    d_in->Resize(dims);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+namespace ops = paddle::operators;
+
+REGISTER_OP(reshape, ops::ReshapeOp, ops::ReshapeOpMaker, reshape_grad,
+            ops::ReshapeGradOp);
+REGISTER_OP_CPU_KERNEL(reshape,
+                       ops::ReshapeKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    reshape_grad, ops::ReshapeGradKernel<paddle::platform::CPUPlace, float>);
diff --git a/paddle/operators/reshape_op.cu b/paddle/operators/reshape_op.cu
new file mode 100644
index 0000000000..23dbe089d3
--- /dev/null
+++ b/paddle/operators/reshape_op.cu
@@ -0,0 +1,22 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/operators/reshape_op.h"
+
+REGISTER_OP_GPU_KERNEL(
+    reshape,
+    paddle::operators::ReshapeKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    reshape_grad,
+    paddle::operators::ReshapeGradKernel<paddle::platform::GPUPlace, float>);
diff --git a/paddle/operators/reshape_op.h b/paddle/operators/reshape_op.h
new file mode 100644
index 0000000000..873acf3078
--- /dev/null
+++ b/paddle/operators/reshape_op.h
@@ -0,0 +1,55 @@
+/* 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 "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename Place, typename T>
+class ReshapeKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const {
+    auto* out = ctx.Output<framework::Tensor>("Out");
+    auto* in = ctx.Input<framework::Tensor>("X");
+    out->mutable_data<T>(ctx.GetPlace());
+
+    auto shape = ctx.Attr<std::vector<int>>("shape");
+    std::vector<int64_t> shape_int64(shape.size(), 0);
+    std::transform(shape.begin(), shape.end(), shape_int64.begin(),
+                   [](int a) { return static_cast<int64_t>(a); });
+    auto out_dims = framework::make_ddim(shape_int64);
+    out->CopyFrom<T>(*in, ctx.GetPlace());
+    out->Resize(out_dims);
+  }
+};
+
+template <typename Place, typename T>
+class ReshapeGradKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const {
+    auto* d_out = ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
+    auto* d_x = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
+    d_x->mutable_data<T>(ctx.GetPlace());
+
+    auto in_dims = d_x->dims();
+    d_x->CopyFrom<T>(*d_out, ctx.GetPlace());
+    d_x->Resize(in_dims);
+  }
+};
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/operators/rnn/recurrent_op_utils.cc b/paddle/operators/rnn/recurrent_op_utils.cc
index 97872c67ac..6c082cb182 100644
--- a/paddle/operators/rnn/recurrent_op_utils.cc
+++ b/paddle/operators/rnn/recurrent_op_utils.cc
@@ -21,6 +21,7 @@ namespace rnn {
 namespace f = paddle::framework;
 
 using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
 
 void SegmentInputs(const std::vector<Scope*>& step_scopes,
                    const std::vector<Link>& inlinks, const size_t seq_len,
@@ -31,7 +32,7 @@ void SegmentInputs(const std::vector<Scope*>& step_scopes,
     PADDLE_ENFORCE(input_var != nullptr, "input link [%s] is not in scope.",
                    inlinks[i].external);
 
-    Tensor* input = input_var->GetMutable<Tensor>();
+    LoDTensor* input = input_var->GetMutable<LoDTensor>();
     f::DDim dims = input->dims();
     PADDLE_ENFORCE(static_cast<size_t>(dims[0]) == seq_len,
                    "all the inlinks must have same length");
@@ -40,6 +41,8 @@ void SegmentInputs(const std::vector<Scope*>& step_scopes,
       Tensor* step_input =
           step_scopes[j]->NewVar(inlinks[i].internal)->GetMutable<Tensor>();
       if (!infer_shape_mode) {
+        // The input of operators of each step is Tensor here.
+        // Maybe need to modify Slice function.
         *step_input = input->Slice<float>(j, j + 1);
       }
       step_input->Resize(step_dims);
@@ -54,21 +57,23 @@ void ConcatOutputs(const std::vector<Scope*>& step_scopes,
     auto output_var = step_scopes[0]->FindVar(outlinks[i].external);
     PADDLE_ENFORCE(output_var != nullptr, "output link [%s] is not in scope.",
                    outlinks[i].external);
-    Tensor* output = output_var->GetMutable<Tensor>();
+    LoDTensor* output = output_var->GetMutable<LoDTensor>();
 
     if (infer_shape_mode) {
       auto step_scope_var = step_scopes[0]->FindVar(outlinks[i].internal);
       PADDLE_ENFORCE(step_scope_var != nullptr, "%s not in scope",
                      outlinks[i].internal);
-      f::DDim step_dims = step_scope_var->template GetMutable<Tensor>()->dims();
+      f::DDim step_dims =
+          step_scope_var->template GetMutable<LoDTensor>()->dims();
       std::vector<int64_t> dims_vec = vectorize(step_dims);
       dims_vec.insert(dims_vec.begin(), seq_len);
       output->Resize(f::make_ddim(dims_vec));
     } else {
       output->mutable_data<float>(platform::CPUPlace());
       for (size_t j = 0; j < seq_len; j++) {
-        Tensor* step_output =
-            step_scopes[j]->FindVar(outlinks[i].internal)->GetMutable<Tensor>();
+        LoDTensor* step_output = step_scopes[j]
+                                     ->FindVar(outlinks[i].internal)
+                                     ->GetMutable<LoDTensor>();
         // TODO(luotao02) data type and platform::DeviceContext() should set
         // correctly
         (output->Slice<float>(j, j + 1))
@@ -94,8 +99,8 @@ void LinkMemories(const std::vector<Scope*>& scopes,
   auto scope = scopes[step_id];
   auto linked_scope = scopes[step_id + offset];
   for (auto& attr : memories) {
-    auto mem = scope->FindVar(attr.pre_var)->GetMutable<Tensor>();
-    auto linked_mem = linked_scope->FindVar(attr.var)->GetMutable<Tensor>();
+    auto mem = scope->FindVar(attr.pre_var)->GetMutable<LoDTensor>();
+    auto linked_mem = linked_scope->FindVar(attr.var)->GetMutable<LoDTensor>();
     if (infer_shape_mode) {
       mem->Resize(linked_mem->dims());
     } else {
diff --git a/paddle/operators/rowwise_add_op.cc b/paddle/operators/rowwise_add_op.cc
index fa8f0ff1a8..2a3fd3be94 100644
--- a/paddle/operators/rowwise_add_op.cc
+++ b/paddle/operators/rowwise_add_op.cc
@@ -25,6 +25,13 @@ class RowwiseAddOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of RowwiseAddOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("b"),
+                            "Input(b) of RowwiseAddOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of RowwiseAddOp should not be null.");
+
     auto x_dims = ctx.Input<Tensor>("X")->dims();
     auto b_dims = ctx.Input<Tensor>("b")->dims();
     PADDLE_ENFORCE_GT(
@@ -37,7 +44,7 @@ class RowwiseAddOp : public framework::OperatorWithKernel {
         framework::slice_ddim(x_dims, num_col_dims, x_dims.size()), b_dims,
         "The width of two operands must be same");
     PADDLE_ENFORCE_EQ(ctx.OutputSize("Out"), 1, "The output size must be 1");
-    ctx.Output<Tensor>("Out")->Resize(x_dims);
+    ctx.Output<framework::LoDTensor>("Out")->Resize(x_dims);
   }
 };
 
@@ -76,8 +83,8 @@ class RowwiseAddGradOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(
         framework::slice_ddim(x_dims, num_col_dims, x_dims.size()), b_dims,
         "The width of two operands must be same");
-    auto *dx = ctx.Output<Tensor>(framework::GradVarName("X"));
-    auto *db = ctx.Output<Tensor>(framework::GradVarName("b"));
+    auto *dx = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto *db = ctx.Output<framework::LoDTensor>(framework::GradVarName("b"));
     if (dx) dx->Resize(x_dims);
     if (db) db->Resize(b_dims);
   }
diff --git a/paddle/operators/scale_op.cc b/paddle/operators/scale_op.cc
index ea991f683d..d1f42e8662 100644
--- a/paddle/operators/scale_op.cc
+++ b/paddle/operators/scale_op.cc
@@ -27,8 +27,13 @@ class ScaleOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of ScaleOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of ScaleOp should not be null.");
+
     auto *in = ctx.Input<framework::Tensor>("X");
-    auto *out = ctx.Output<framework::Tensor>("Out");
+    auto *out = ctx.Output<framework::LoDTensor>("Out");
     out->Resize(in->dims());
   }
 };
diff --git a/paddle/operators/scatter_op.cc b/paddle/operators/scatter_op.cc
index f901edefa2..8820262732 100644
--- a/paddle/operators/scatter_op.cc
+++ b/paddle/operators/scatter_op.cc
@@ -24,6 +24,15 @@ class ScatterOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Ref"),
+                            "Input(Ref) of ScatterOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Index"),
+                            "Input(Index) of ScatterOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Updates"),
+                            "Input(Updates) of ScatterOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of ScatterOp should not be null.");
+
     PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("Index")->dims().size(), 1,
                       "Update Index should be 1-D.");
     PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("Ref")->dims().size(),
@@ -35,7 +44,8 @@ class ScatterOp : public framework::OperatorWithKernel {
     framework::DDim data_dim(ctx.Input<Tensor>("Updates")->dims());
     for (int i = 1; i < data_dim.size(); ++i)
       PADDLE_ENFORCE_EQ(data_dim[i], ctx.Input<Tensor>("Updates")->dims()[i]);
-    ctx.Output<Tensor>("Out")->Resize(ctx.Input<Tensor>("Ref")->dims());
+    ctx.Output<framework::LoDTensor>("Out")->Resize(
+        ctx.Input<Tensor>("Ref")->dims());
   }
 };
 
@@ -45,9 +55,11 @@ class ScatterGradOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    auto *dUpdates = ctx.Output<Tensor>(framework::GradVarName("Updates"));
+    auto *dUpdates =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("Updates"));
     auto *Updates = ctx.Input<Tensor>("Updates");
-    auto *dRef = ctx.Output<Tensor>(framework::GradVarName("Ref"));
+    auto *dRef =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("Ref"));
     auto *Ref = ctx.Input<Tensor>("Ref");
 
     dRef->Resize(Ref->dims());
diff --git a/paddle/operators/sequence_avg_pool_op.cc b/paddle/operators/sequence_avg_pool_op.cc
new file mode 100644
index 0000000000..9815b8f3a8
--- /dev/null
+++ b/paddle/operators/sequence_avg_pool_op.cc
@@ -0,0 +1,95 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/sequence_avg_pool_op.h"
+
+namespace paddle {
+namespace operators {
+
+class SequenceAvgPoolOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext& ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.InputVar("X"), "Input(X) of SequenceAvgPoolOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Out"),
+        "Output(Out) of SequenceAvgPoolOp should not be null.");
+
+    auto* x = ctx.Input<framework::LoDTensor>("X");
+    auto dims = x->dims();
+    auto lod = x->lod();
+    PADDLE_ENFORCE_EQ(lod.size(), 1UL, "Only support one level sequence now.");
+    PADDLE_ENFORCE_GE(
+        dims[0],
+        /*batch size = */ static_cast<int64_t>(lod[0].size() - 1),
+        "The first dimension of Input(X) must be large than batch size.");
+    dims[0] = lod[0].size() - 1;
+    ctx.Output<framework::LoDTensor>("Out")->Resize({dims});
+  }
+};
+
+class SequenceAvgPoolOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  SequenceAvgPoolOpMaker(framework::OpProto* proto,
+                         framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "Input of SequenceAvgPoolOp.");
+    AddOutput("Out", "The output of SequenceAvgPoolOp.");
+    AddComment(R"DOC(
+    SequenceAvgPoolOp averages features of all time-steps of each instance.
+    More detailed comments will be added later.
+    )DOC");
+  }
+};
+
+class SequenceAvgPoolGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext& ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
+                            "Gradient of Out should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "The input X should not be null.");
+    auto og_dims =
+        ctx.Input<framework::LoDTensor>(framework::GradVarName("Out"))->dims();
+    auto x_dims = ctx.Input<framework::LoDTensor>("X")->dims();
+    PADDLE_ENFORCE_EQ(og_dims.size(), x_dims.size(),
+                      "The rank of output grad must equal to Input(X).");
+    for (int64_t i = 1; i < og_dims.size(); ++i) {
+      PADDLE_ENFORCE_EQ(og_dims[i], x_dims[i], "The dimension mismatch.");
+    }
+    auto* x_grad =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    x_grad->Resize(x_dims);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(sequence_avg_pool, ops::SequenceAvgPoolOp,
+            ops::SequenceAvgPoolOpMaker, sequence_avg_pool_grad,
+            ops::SequenceAvgPoolGradOp);
+REGISTER_OP_CPU_KERNEL(
+    sequence_avg_pool,
+    ops::SequenceAvgPoolKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    sequence_avg_pool_grad,
+    ops::SequenceAvgPoolGradKernel<paddle::platform::CPUPlace, float>);
diff --git a/paddle/operators/sequence_avg_pool_op.cu b/paddle/operators/sequence_avg_pool_op.cu
new file mode 100644
index 0000000000..bc9d1611fc
--- /dev/null
+++ b/paddle/operators/sequence_avg_pool_op.cu
@@ -0,0 +1,25 @@
+/* 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. */
+
+#define EIGEN_USE_GPU
+
+#include "paddle/operators/sequence_avg_pool_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(
+    sequence_avg_pool,
+    ops::SequenceAvgPoolKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    sequence_avg_pool_grad,
+    ops::SequenceAvgPoolGradKernel<paddle::platform::GPUPlace, float>);
diff --git a/paddle/operators/sequence_avg_pool_op.h b/paddle/operators/sequence_avg_pool_op.h
new file mode 100644
index 0000000000..ebe0956344
--- /dev/null
+++ b/paddle/operators/sequence_avg_pool_op.h
@@ -0,0 +1,84 @@
+/* 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 "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
+
+template <typename Place, typename T>
+class SequenceAvgPoolKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* in = context.Input<LoDTensor>("X");
+    auto* out = context.Output<LoDTensor>("Out");
+
+    auto dims = in->dims();
+    auto lod = in->lod();
+    int64_t w = in->numel() / dims[0];
+
+    out->mutable_data<T>(context.GetPlace());
+    auto place = context.GetEigenDevice<Place>();
+    for (int i = 0; i < static_cast<int>(lod[0].size()) - 1; ++i) {
+      Tensor in_t = in->Slice<T>(static_cast<int>(lod[0][i]),
+                                 static_cast<int>(lod[0][i + 1]));
+      Tensor out_t = out->Slice<T>(i, i + 1);
+      int64_t h = static_cast<int64_t>(lod[0][i + 1] - lod[0][i]);
+      auto in_e = EigenMatrix<T>::From(in_t, framework::make_ddim({h, w}));
+      auto out_e = EigenVector<T>::Flatten(out_t);
+      out_e.device(place) = in_e.mean(Eigen::array<int, 1>({{0}}));
+    }
+  }
+};
+
+template <typename Place, typename T>
+class SequenceAvgPoolGradKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* in = context.Input<LoDTensor>("X");
+    auto* out_g = context.Input<LoDTensor>(framework::GradVarName("Out"));
+    auto* in_g = context.Output<LoDTensor>(framework::GradVarName("X"));
+
+    auto dims = in->dims();
+    auto lod = in->lod();
+    int64_t w = in->numel() / dims[0];
+
+    in_g->mutable_data<T>(context.GetPlace());
+    auto place = context.GetEigenDevice<Place>();
+    for (int i = 0; i < static_cast<int>(lod[0].size()) - 1; ++i) {
+      auto in_g_t = in_g->Slice<T>(static_cast<int>(lod[0][i]),
+                                   static_cast<int>(lod[0][i + 1]));
+      auto out_g_t = out_g->Slice<T>(i, i + 1);
+      int64_t h = static_cast<int64_t>(lod[0][i + 1] - lod[0][i]);
+      auto in_g_e = EigenMatrix<T>::From(in_g_t, {h, w});
+      auto out_g_e = EigenMatrix<T>::From(out_g_t, {1, w});
+      Eigen::DSizes<int, 2> bcast(h, 1);
+      in_g_e.device(place) = (out_g_e / static_cast<T>(h)).broadcast(bcast);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/operators/sgd_op.cc b/paddle/operators/sgd_op.cc
index ad267e7f08..1232e64c7f 100644
--- a/paddle/operators/sgd_op.cc
+++ b/paddle/operators/sgd_op.cc
@@ -23,10 +23,18 @@ class SGDOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(
-        ctx.Input<Tensor>("param")->dims() == ctx.Input<Tensor>("grad")->dims(),
-        "Two input of SGD Op's dimension must be same.");
-    ctx.Output<Tensor>("param_out")->Resize(ctx.Input<Tensor>("param")->dims());
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("param"),
+                            "Input(param) of SGDOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("grad"),
+                            "Input(grad) of SGDOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("param_out"),
+                            "Output(param_out) of SGDOp should not be null.");
+
+    PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("param")->dims(),
+                      ctx.Input<Tensor>("grad")->dims(),
+                      "Two input of SGD Op's dimension must be same.");
+    ctx.Output<framework::LoDTensor>("param_out")
+        ->Resize(ctx.Input<Tensor>("param")->dims());
   }
 };
 
diff --git a/paddle/operators/sigmoid_op.cc b/paddle/operators/sigmoid_op.cc
index 761c6de8d4..992b19965e 100644
--- a/paddle/operators/sigmoid_op.cc
+++ b/paddle/operators/sigmoid_op.cc
@@ -23,7 +23,13 @@ class SigmoidOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    ctx.Output<Tensor>("Y")->Resize(ctx.Input<Tensor>("X")->dims());
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of SigmoidOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Y"),
+                            "Output(Y) of SigmoidOp should not be null.");
+
+    ctx.Output<framework::LoDTensor>("Y")->Resize(
+        ctx.Input<Tensor>("X")->dims());
   }
 };
 
@@ -44,7 +50,7 @@ class SigmoidOpGrad : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    ctx.Output<Tensor>(framework::GradVarName("X"))
+    ctx.Output<framework::LoDTensor>(framework::GradVarName("X"))
         ->Resize(ctx.Input<Tensor>("Y")->dims());
   }
 };
diff --git a/paddle/operators/softmax_op.cc b/paddle/operators/softmax_op.cc
index 7166b2f60b..c67eb028c8 100644
--- a/paddle/operators/softmax_op.cc
+++ b/paddle/operators/softmax_op.cc
@@ -23,9 +23,15 @@ class SoftmaxOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of SoftmaxOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Y"),
+                            "Output(Y) of SoftmaxOp should not be null.");
+
     PADDLE_ENFORCE(ctx.Input<Tensor>("X")->dims().size() == 2UL,
                    "The input of softmax op must be a matrix.");
-    ctx.Output<Tensor>("Y")->Resize(ctx.Input<Tensor>("X")->dims());
+    ctx.Output<framework::LoDTensor>("Y")->Resize(
+        ctx.Input<Tensor>("X")->dims());
   }
 };
 
@@ -71,7 +77,7 @@ class SoftmaxOpGrad : public framework::OperatorWithKernel {
                       ctx.Input<Tensor>(framework::GradVarName("Y"))->dims(),
                       "Input(Y) and its gradients should have a same shape.");
 
-    ctx.Output<Tensor>(framework::GradVarName("X"))
+    ctx.Output<framework::LoDTensor>(framework::GradVarName("X"))
         ->Resize(ctx.Input<Tensor>("X")->dims());
   }
 };
diff --git a/paddle/operators/split_op.cc b/paddle/operators/split_op.cc
new file mode 100644
index 0000000000..61296f5c81
--- /dev/null
+++ b/paddle/operators/split_op.cc
@@ -0,0 +1,118 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/split_op.h"
+#include "paddle/operators/net_op.h"
+
+namespace paddle {
+namespace operators {
+using framework::Tensor;
+
+class SplitOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    // infershape
+    auto *in = ctx.Input<framework::Tensor>("X");
+    auto outs = ctx.MultiOutput<framework::LoDTensor>("Out");
+    size_t axis = static_cast<size_t>(ctx.Attr<int>("axis"));
+    size_t num = static_cast<size_t>(ctx.Attr<int>("num"));
+    std::vector<int> sections =
+        static_cast<std::vector<int>>(ctx.Attr<std::vector<int>>("sections"));
+    const size_t n = outs.size();
+
+    if (num > 0) {
+      int64_t in_axis_dim = in->dims()[axis];
+      PADDLE_ENFORCE_EQ(in_axis_dim % num, 0,
+                        "tensor split does not result"
+                        " in an equal division");
+      size_t out_axis_dim = in_axis_dim / num;
+      for (size_t i = 0; i < n; ++i) {
+        auto dim = in->dims();
+        dim[axis] = out_axis_dim;
+        outs[i]->Resize(dim);
+      }
+    } else if (sections.size() > 0) {
+      PADDLE_ENFORCE_EQ(sections.size(), n,
+                        "tensor split sections size"
+                        "should be equal to output size.");
+      for (size_t i = 0; i < n; ++i) {
+        auto dim = in->dims();
+        dim[axis] = sections[i];
+        outs[i]->Resize(dim);
+      }
+    } else {
+      PADDLE_ENFORCE_NOT_NULL(nullptr, "split operator should",
+                              " specify indices or sections.");
+    }
+  }
+};
+
+class SplitOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  SplitOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "the input tensor of split operator.");
+    AddOutput("Out", "the output tensors of split operator.").AsDuplicable();
+    AddComment(R"DOC(
+      Split the input tensor into multiple sub-tensors.
+      Example:
+        Input = [[1,2],
+                 [3,4],
+                 [5,6]]
+        sections = [2,1]
+        axis = 0
+        Output[0] = [[1,2],
+                     [3,4]]
+        Output[1] = [[5,6]]
+
+    )DOC");
+    AddAttr<std::vector<int>>("sections",
+                              "the length for each"
+                              "output along with the specify axis.")
+        .SetDefault(std::vector<int>{});
+    AddAttr<int>("num",
+                 "number of the sub-tensors, it must evenly divide "
+                 "Input.dims()[axis]")
+        .SetDefault(0);
+    AddAttr<int>("axis", "The axis which the input will be splited on.")
+        .SetDefault(0);
+  }
+};
+
+class SplitOpGrad : public NetOp {
+ public:
+  SplitOpGrad(const std::string &type, const framework::VariableNameMap &inputs,
+              const framework::VariableNameMap &outputs,
+              const framework::AttributeMap &attrs)
+      : NetOp(type, inputs, outputs, attrs) {
+    auto out_grad = Inputs(framework::GradVarName("Out"));
+    auto x_grad = Output(framework::GradVarName("X"));
+    AppendOp(framework::OpRegistry::CreateOp("concat", {{"X", out_grad}},
+                                             {{"Out", {x_grad}}}, attrs));
+    CompleteAddOp(false);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+USE_CPU_ONLY_OP(concat);
+REGISTER_OP(split, ops::SplitOp, ops::SplitOpMaker, split_grad,
+            ops::SplitOpGrad);
+REGISTER_OP_CPU_KERNEL(split,
+                       ops::SplitKernel<paddle::platform::CPUPlace, float>);
diff --git a/paddle/operators/split_op.h b/paddle/operators/split_op.h
new file mode 100644
index 0000000000..860690ee89
--- /dev/null
+++ b/paddle/operators/split_op.h
@@ -0,0 +1,62 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include <vector>
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename Place, typename T>
+class SplitKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* in = ctx.Input<framework::Tensor>("X");
+    auto outs = ctx.MultiOutput<framework::Tensor>("Out");
+    int64_t axis = static_cast<int64_t>(ctx.Attr<int>("axis"));
+    size_t before = 1, after = 1;
+    const size_t n = outs.size();
+    size_t input_axis_dim = in->dims()[axis];
+
+    for (int64_t i = 0; i < in->dims().size(); ++i) {
+      if (i == axis) {
+        continue;
+      }
+      if (i < axis) {
+        before *= in->dims()[i];
+      } else {
+        after *= in->dims()[i];
+      }
+    }
+    size_t input_offset = 0;
+    for (size_t i = 0; i < n; i++) {
+      auto& out = outs[i];
+      size_t axis_dim = out->dims()[axis];
+      for (size_t j = 0; j < before; j++) {
+        size_t len = axis_dim * after * sizeof(T);
+        T* dest =
+            out->mutable_data<T>(platform::CPUPlace()) + axis_dim * after * j;
+        const T* src =
+            in->data<T>() + input_offset + input_axis_dim * after * j;
+        memcpy(dest, src, len);
+      }
+      input_offset += axis_dim * after;
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/operators/squared_l2_distance_op.cc b/paddle/operators/squared_l2_distance_op.cc
index dc30644a5e..39f4305877 100644
--- a/paddle/operators/squared_l2_distance_op.cc
+++ b/paddle/operators/squared_l2_distance_op.cc
@@ -23,12 +23,18 @@ class SquaredL2DistanceOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext& ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
-                            "Input of SquaredL2DistanceOp "
-                            "must be initialized.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
-                            "Target of SquaredL2DistanceOp "
-                            "must be initialized.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.InputVar("X"),
+        "Input(X) of SquaredL2DistanceOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.InputVar("Y"),
+        "Input(Y) of SquaredL2DistanceOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("sub_result"),
+        "Output(sub_result) of SquaredL2DistanceOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Out"),
+        "Output(Out) of SquaredL2DistanceOp should not be null.");
 
     auto* x = ctx.Input<Tensor>("X");
     auto x_dims = x->dims();
@@ -41,18 +47,16 @@ class SquaredL2DistanceOp : public framework::OperatorWithKernel {
 
     int rank = framework::arity(x_dims);
     PADDLE_ENFORCE_GE(rank, 2, "Tensor rank should be at least equal to 2.");
-    PADDLE_ENFORCE_EQ(framework::product(x_dims) / x_dims[0],
-                      framework::product(y_dims) / y_dims[0],
+    PADDLE_ENFORCE_EQ(x->numel() / x_dims[0], y->numel() / y_dims[0],
                       "Product of dimensions expcet the first dimension of "
                       "input and target must be equal.");
     PADDLE_ENFORCE(y_dims[0] == 1 || y_dims[0] == x_dims[0],
                    "First dimension of target must be equal to input "
                    "or to 1.");
 
-    ctx.Output<Tensor>("sub_result")
-        ->Resize({static_cast<int>(x_dims[0]),
-                  static_cast<int>(framework::product(x_dims) / x_dims[0])});
-    ctx.Output<Tensor>("Out")->Resize({x_dims[0], 1});
+    ctx.Output<framework::LoDTensor>("sub_result")
+        ->Resize({x_dims[0], x->numel() / x_dims[0]});
+    ctx.Output<framework::LoDTensor>("Out")->Resize({x_dims[0], 1});
   }
 };
 
@@ -96,8 +100,10 @@ class SquaredL2DistanceGradOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(out_dims[1], 1,
                       "Second dimension of output gradient "
                       "must be 1.");
-    auto* x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
-    auto* y_grad = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    auto* x_grad =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto* y_grad =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("Y"));
     if (x_grad) x_grad->Resize(x_dims);
     if (y_grad) y_grad->Resize(y_dims);
   }
diff --git a/paddle/operators/squared_l2_distance_op.h b/paddle/operators/squared_l2_distance_op.h
index ad3347a0b3..097ac04fc0 100644
--- a/paddle/operators/squared_l2_distance_op.h
+++ b/paddle/operators/squared_l2_distance_op.h
@@ -39,7 +39,7 @@ class SquaredL2DistanceKernel : public framework::OpKernel {
     auto in0_dims = in0->dims();
     auto in1_dims = in1->dims();
 
-    int cols = framework::product(in0_dims) / in0_dims[0];
+    int cols = in0->numel() / in0_dims[0];
     // reduce dimensions except the first
     auto x =
         EigenMatrix<T>::From(*in0, framework::make_ddim({in0_dims[0], cols}));
@@ -82,7 +82,7 @@ class SquaredL2DistanceGradKernel : public framework::OpKernel {
     auto x_dims = x_g->dims();
     auto y_dims = y_g->dims();
 
-    int cols = framework::product(x_dims) / x_dims[0];
+    int cols = x_g->numel() / x_dims[0];
     // calculate gradient
     auto grad_mat = 2 *
                     (out_grad.broadcast(Eigen::array<int, 2>({{1, cols}}))) *
diff --git a/paddle/operators/sum_op.cc b/paddle/operators/sum_op.cc
new file mode 100644
index 0000000000..41e05c27f9
--- /dev/null
+++ b/paddle/operators/sum_op.cc
@@ -0,0 +1,79 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+http://www.apache.org/licenses/LICENSE-2.0
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/sum_op.h"
+#include <vector>
+
+namespace paddle {
+namespace operators {
+using framework::Tensor;
+
+class SumOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE(!ctx.MultiInputVar("X").empty(),
+                   "Input(X) of SumOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of SumOp should not be null.");
+
+    auto ins = ctx.MultiInput<framework::Tensor>("X");
+    auto *out = ctx.Output<framework::LoDTensor>("Out");
+    int N = ins.size();
+
+    auto in_dim = ins[0]->dims();
+
+    PADDLE_ENFORCE_GT(N, 1, "Input tensors count should > 1.");
+    for (int i = 1; i < N; i++) {
+      auto dim = ins[i]->dims();
+      PADDLE_ENFORCE(in_dim == dim, "Input tensors must have same shape");
+    }
+    out->Resize(in_dim);
+  }
+};
+
+class SumOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  SumOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "the input tensors of sum operator.").AsDuplicable();
+    AddOutput("Out", "the output tensor of sum operator.");
+    AddComment(R"DOC(
+            Sum the input tensors.
+        )DOC");
+  }
+};
+
+class SumGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    auto outputs =
+        ctx.MultiOutput<framework::LoDTensor>(framework::GradVarName("X"));
+    auto dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
+    for (auto output : outputs) {
+      output->Resize(dims);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(sum, ops::SumOp, ops::SumOpMaker, sum_grad, ops::SumGradOp);
+REGISTER_OP_CPU_KERNEL(sum, ops::SumKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(sum_grad,
+                       ops::SumGradKernel<paddle::platform::CPUPlace, float>);
diff --git a/paddle/operators/sum_op.cu b/paddle/operators/sum_op.cu
new file mode 100644
index 0000000000..a465cf3659
--- /dev/null
+++ b/paddle/operators/sum_op.cu
@@ -0,0 +1,18 @@
+/* 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. */
+
+#define EIGEN_USE_GPU
+#include "paddle/operators/sum_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(sum, ops::SumKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(sum_grad,
+                       ops::SumGradKernel<paddle::platform::GPUPlace, float>);
diff --git a/paddle/operators/sum_op.h b/paddle/operators/sum_op.h
new file mode 100644
index 0000000000..0b1e9ebaa3
--- /dev/null
+++ b/paddle/operators/sum_op.h
@@ -0,0 +1,65 @@
+/* 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 "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
+
+template <typename Place, typename T>
+class SumKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto ins = context.MultiInput<Tensor>("X");
+    auto* out = context.Output<Tensor>("Out");
+    out->mutable_data<T>(context.GetPlace());
+
+    auto place = context.GetEigenDevice<Place>();
+    auto result = EigenVector<T>::Flatten(*out);
+
+    int N = ins.size();
+    auto in = EigenVector<T>::Flatten(*(ins[0]));
+    result.device(place) = in;
+    for (int i = 1; i < N; i++) {
+      auto in = EigenVector<T>::Flatten(*(ins[i]));
+      result.device(place) = result + in;
+    }
+  }
+};
+
+template <typename Place, typename T>
+class SumGradKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* input = context.Input<Tensor>(framework::GradVarName("Out"));
+    auto outs = context.MultiOutput<Tensor>(framework::GradVarName("X"));
+    for (auto out : outs) {
+      out->mutable_data<T>(context.GetPlace());
+    }
+
+    auto place = context.GetEigenDevice<Place>();
+    auto in = EigenVector<T>::Flatten(*input);
+    for (auto out : outs) {
+      auto result = EigenVector<T>::Flatten(*out);
+      result.device(place) = in;
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/operators/top_k_op.cc b/paddle/operators/top_k_op.cc
index 38d2f0a09a..169b815fef 100644
--- a/paddle/operators/top_k_op.cc
+++ b/paddle/operators/top_k_op.cc
@@ -24,7 +24,12 @@ class TopkOp : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
-                            "Input of TopkOP must be initialized.");
+                            "Input(X) of TopkOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of TopkOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Indices"),
+                            "Output(Indices) of TopkOp should not be null.");
+
     auto *input = ctx.Input<framework::Tensor>("X");
     const int k = static_cast<int>(ctx.Attr<int>("k"));
 
@@ -35,8 +40,8 @@ class TopkOp : public framework::OperatorWithKernel {
 
     framework::DDim dims = input->dims();
     dims[dims.size() - 1] = k;
-    ctx.Output<Tensor>("Out")->Resize(dims);
-    ctx.Output<Tensor>("Indices")->Resize(dims);
+    ctx.Output<framework::LoDTensor>("Out")->Resize(dims);
+    ctx.Output<framework::LoDTensor>("Indices")->Resize(dims);
   }
 };
 
diff --git a/paddle/operators/uniform_random_op.cc b/paddle/operators/uniform_random_op.cc
index f2aeef6c31..184bcbc29c 100644
--- a/paddle/operators/uniform_random_op.cc
+++ b/paddle/operators/uniform_random_op.cc
@@ -35,7 +35,7 @@ class CPUUniformRandomKernel : public framework::OpKernel {
     std::uniform_real_distribution<T> dist(
         static_cast<T>(context.Attr<float>("min")),
         static_cast<T>(context.Attr<float>("max")));
-    int64_t size = framework::product(tensor->dims());
+    int64_t size = tensor->numel();
     for (int64_t i = 0; i < size; ++i) {
       data[i] = dist(engine);
     }
@@ -48,9 +48,13 @@ class UniformRandomOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext& ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Out"),
+        "Output(Out) of UniformRandomOp should not be null.");
+
     PADDLE_ENFORCE(Attr<float>("min") < Attr<float>("max"),
                    "uniform_random's min must less then max");
-    auto* tensor = ctx.Output<framework::Tensor>("Out");
+    auto* tensor = ctx.Output<framework::LoDTensor>("Out");
     auto dims = Attr<std::vector<int>>("dims");
     std::vector<int64_t> temp;
     temp.reserve(dims.size());
diff --git a/paddle/operators/uniform_random_op.cu b/paddle/operators/uniform_random_op.cu
index c2c041b144..6614b53b3f 100644
--- a/paddle/operators/uniform_random_op.cu
+++ b/paddle/operators/uniform_random_op.cu
@@ -53,8 +53,8 @@ class GPUUniformRandomKernel : public framework::OpKernel {
     T min = static_cast<T>(context.Attr<float>("min"));
     T max = static_cast<T>(context.Attr<float>("max"));
     thrust::counting_iterator<unsigned int> index_sequence_begin(0);
-    ssize_t N = framework::product(tensor->dims());
-    thrust::transform(index_sequence_begin, index_sequence_begin + N,
+    int64_t size = tensor->numel();
+    thrust::transform(index_sequence_begin, index_sequence_begin + size,
                       thrust::device_ptr<T>(data),
                       UniformGenerator<T>(min, max, seed));
   }
diff --git a/paddle/platform/CMakeLists.txt b/paddle/platform/CMakeLists.txt
index 17bdac8749..8b605e51c3 100644
--- a/paddle/platform/CMakeLists.txt
+++ b/paddle/platform/CMakeLists.txt
@@ -24,3 +24,4 @@ cc_library(device_context SRCS device_context.cc DEPS memory buddy_allocator
 nv_test(device_context_test SRCS device_context_test.cc DEPS device_context gpu_info)
 
 nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda)
+nv_test(transform_test SRCS transform_test.cu DEPS paddle_memory place)
diff --git a/paddle/platform/cuda_helper.h b/paddle/platform/cuda_helper.h
index 6feec0d7f8..a7d99cde10 100644
--- a/paddle/platform/cuda_helper.h
+++ b/paddle/platform/cuda_helper.h
@@ -24,6 +24,11 @@ namespace platform {
 #define USE_CUDA_ATOMIC(op, T) \
   CUDA_ATOMIC_WRAPPER(op, T) { return atomic##op(address, val); }
 
+// Default thread count per block(or block size).
+// TODO(typhoonzero): need to benchmark against setting this value
+//                    to 1024.
+constexpr int PADDLE_CUDA_NUM_THREADS = 512;
+
 // For atomicAdd.
 USE_CUDA_ATOMIC(Add, float);
 
diff --git a/paddle/platform/details/device_ptr_cast.h b/paddle/platform/details/device_ptr_cast.h
new file mode 100644
index 0000000000..4015491fcd
--- /dev/null
+++ b/paddle/platform/details/device_ptr_cast.h
@@ -0,0 +1,56 @@
+/* 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
+
+#ifndef __NVCC__
+#error device_ptr_cast must be include by .cu file
+#endif
+
+#include <thrust/device_ptr.h>
+
+namespace paddle {
+namespace platform {
+namespace details {
+template <typename T, bool is_ptr>
+struct DevicePtrCast;
+
+template <typename T>
+struct DevicePtrCast<T, true> {
+  using ELEM = typename std::remove_pointer<T>::type;
+  using RTYPE = thrust::device_ptr<ELEM>;
+
+  inline thrust::device_ptr<ELEM> operator()(ELEM* ele) const {
+    return thrust::device_pointer_cast(ele);
+  }
+};
+
+template <typename T>
+struct DevicePtrCast<T, false> {
+  using RTYPE = T;
+  inline RTYPE operator()(RTYPE it) const { return it; }
+};
+
+// Cast T to thrust::device_ptr if T is a pointer.
+// Otherwise, e.g., T is a iterator, return T itself.
+template <typename T>
+auto DevPtrCast(T t) ->
+    typename DevicePtrCast<T, std::is_pointer<T>::value>::RTYPE {
+  DevicePtrCast<T, std::is_pointer<T>::value> cast;
+  return cast(t);
+}
+
+}  // namespace details
+}  // namespace platform
+}  // namespace paddle
diff --git a/paddle/platform/enforce.h b/paddle/platform/enforce.h
index 81448897e9..df5f71ed76 100644
--- a/paddle/platform/enforce.h
+++ b/paddle/platform/enforce.h
@@ -78,7 +78,7 @@ struct EnforceNotMet : public std::exception {
 
       Dl_info info;
       for (int i = 0; i < size; ++i) {
-        if (dladdr(call_stack[i], &info)) {
+        if (dladdr(call_stack[i], &info) && info.dli_sname) {
           auto demangled = demangle(info.dli_sname);
           auto addr_offset = static_cast<char*>(call_stack[i]) -
                              static_cast<char*>(info.dli_saddr);
diff --git a/paddle/platform/transform.h b/paddle/platform/transform.h
new file mode 100644
index 0000000000..3ee4acd296
--- /dev/null
+++ b/paddle/platform/transform.h
@@ -0,0 +1,66 @@
+/* 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 "paddle/platform/enforce.h"
+#include "paddle/platform/hostdevice.h"
+#include "paddle/platform/place.h"
+
+#include <algorithm>
+#include <type_traits>
+#ifdef __NVCC__
+#include <thrust/transform.h>
+#include "paddle/platform/details/device_ptr_cast.h"
+#endif
+
+namespace paddle {
+namespace platform {
+// Transform on host or device. It provides the same API in std library.
+template <typename Place, typename InputIter, typename OutputIter,
+          typename UnaryOperation>
+void Transform(Place place, InputIter first, InputIter last, OutputIter result,
+               UnaryOperation op) {
+  if (is_cpu_place(place)) {
+    std::transform(first, last, result, op);
+  } else {
+#ifdef __NVCC__
+    using namespace details;
+    thrust::transform(DevPtrCast(first), DevPtrCast(last), DevPtrCast(result),
+                      op);
+#else
+    PADDLE_THROW("Do not invoke `Transform<GPUPlace>` in .cc file");
+#endif
+  }
+}
+
+template <typename Place, typename InputIter1, typename InputIter2,
+          typename OutputIter, typename BinaryOperation>
+void Transform(Place place, InputIter1 first1, InputIter1 last1,
+               InputIter2 first2, OutputIter result, BinaryOperation op) {
+  if (is_cpu_place(place)) {
+    std::transform(first1, last1, first2, result, op);
+  } else {
+#ifdef __NVCC__
+    using namespace details;
+    thrust::transform(DevPtrCast(first1), DevPtrCast(last1), DevPtrCast(first2),
+                      DevPtrCast(result), op);
+#else
+    PADDLE_THROW("Do not invoke `Transform<GPUPlace>` in .cc file");
+#endif
+  }
+};
+
+}  // namespace platform
+}  // namespace paddle
diff --git a/paddle/platform/transform_test.cu b/paddle/platform/transform_test.cu
new file mode 100644
index 0000000000..600fed8f45
--- /dev/null
+++ b/paddle/platform/transform_test.cu
@@ -0,0 +1,84 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include <gtest/gtest.h>
+#include "paddle/memory/memcpy.h"
+#include "paddle/memory/memory.h"
+#include "paddle/platform/transform.h"
+
+template <typename T>
+class Scale {
+ public:
+  explicit Scale(const T& scale) : scale_(scale) {}
+
+  HOSTDEVICE T operator()(const T& a) const { return a * scale_; }
+
+ private:
+  T scale_;
+};
+
+template <typename T>
+class Multiply {
+ public:
+  HOSTDEVICE T operator()(const T& a, const T& b) const { return a * b; }
+};
+
+TEST(Transform, CPUUnary) {
+  using namespace paddle::platform;
+  float buf[4] = {0.1, 0.2, 0.3, 0.4};
+  Transform(CPUPlace(), buf, buf + 4, buf, Scale<float>(10));
+  for (int i = 0; i < 4; ++i) {
+    ASSERT_NEAR(buf[i], static_cast<float>(i + 1), 1e-5);
+  }
+}
+
+TEST(Transform, GPUUnary) {
+  using namespace paddle::platform;
+  using namespace paddle::memory;
+  GPUPlace gpu0(0);
+  float cpu_buf[4] = {0.1, 0.2, 0.3, 0.4};
+  float* gpu_buf = static_cast<float*>(Alloc(gpu0, sizeof(float) * 4));
+  Copy(gpu0, gpu_buf, CPUPlace(), cpu_buf, sizeof(cpu_buf));
+  Transform(gpu0, gpu_buf, gpu_buf + 4, gpu_buf, Scale<float>(10));
+  Copy(CPUPlace(), cpu_buf, gpu0, gpu_buf, sizeof(cpu_buf));
+  Free(gpu0, gpu_buf);
+  for (int i = 0; i < 4; ++i) {
+    ASSERT_NEAR(cpu_buf[i], static_cast<float>(i + 1), 1e-5);
+  }
+}
+
+TEST(Transform, CPUBinary) {
+  using namespace paddle::platform;
+  using namespace paddle::memory;
+  int buf[4] = {1, 2, 3, 4};
+  Transform(CPUPlace(), buf, buf + 4, buf, buf, Multiply<int>());
+  for (int i = 0; i < 4; ++i) {
+    ASSERT_EQ((i + 1) * (i + 1), buf[i]);
+  }
+}
+
+TEST(Transform, GPUBinary) {
+  using namespace paddle::platform;
+  using namespace paddle::memory;
+  int buf[4] = {1, 2, 3, 4};
+  GPUPlace gpu0(0);
+  int* gpu_buf = static_cast<int*>(Alloc(gpu0, sizeof(buf)));
+  Copy(gpu0, gpu_buf, CPUPlace(), buf, sizeof(buf));
+  Transform(gpu0, gpu_buf, gpu_buf + 4, gpu_buf, gpu_buf, Multiply<int>());
+  Copy(CPUPlace(), buf, gpu0, gpu_buf, sizeof(buf));
+  Free(gpu0, gpu_buf);
+  for (int i = 0; i < 4; ++i) {
+    ASSERT_EQ((i + 1) * (i + 1), buf[i]);
+  }
+}
\ No newline at end of file
diff --git a/paddle/pybind/CMakeLists.txt b/paddle/pybind/CMakeLists.txt
index 0003005070..4f05406c7f 100644
--- a/paddle/pybind/CMakeLists.txt
+++ b/paddle/pybind/CMakeLists.txt
@@ -1,5 +1,5 @@
 if(WITH_PYTHON)
-cc_library(paddle_pybind SHARED
+  cc_library(paddle_pybind SHARED
     SRCS pybind.cc
     DEPS pybind python backward
     ${GLOB_OP_LIB})
diff --git a/paddle/pybind/pybind.cc b/paddle/pybind/pybind.cc
index 53985933ed..c7009a604f 100644
--- a/paddle/pybind/pybind.cc
+++ b/paddle/pybind/pybind.cc
@@ -17,11 +17,14 @@ limitations under the License. */
 #include <vector>
 
 #include "paddle/framework/backward.h"
+#include "paddle/framework/lod_tensor.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/operators/cond_op.h"
 #include "paddle/operators/net_op.h"
 #include "paddle/operators/recurrent_op.h"
 #include "paddle/platform/enforce.h"
 #include "paddle/platform/place.h"
+#include "paddle/pybind/pybind.h"
 #include "paddle/pybind/tensor_py.h"
 #include "paddle/string/to_string.h"
 #include "pybind11/numpy.h"
@@ -30,32 +33,12 @@ limitations under the License. */
 
 namespace py = pybind11;
 
-USE_OP(add);
-USE_OP(onehot_cross_entropy);
-USE_OP(sgd);
-USE_OP(mul);
-USE_OP(mean);
-USE_OP(sigmoid);
-USE_OP(softmax);
-USE_OP(rowwise_add);
-USE_OP(fill_zeros_like);
-USE_NO_KERNEL_OP(recurrent);
-USE_OP(gaussian_random);
-USE_OP(uniform_random);
-USE_OP(lookup_table);
-USE_OP(scale);
-USE_NO_KERNEL_OP(identity);
-USE_OP(minus);
-USE_OP(cos_sim);
-USE_CPU_ONLY_OP(gather);
-USE_CPU_ONLY_OP(scatter);
-USE_OP(top_k);
-USE_OP(squared_l2_distance);
-
 namespace paddle {
 namespace framework {
 
 using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
+using LoD = framework::LoD;
 
 static size_t UniqueIntegerGenerator() {
   static std::atomic<size_t> generator;
@@ -115,6 +98,51 @@ PYBIND11_PLUGIN(core) {
         return self.data<float>()[offset];
       });
 
+  py::class_<LoDTensor, Tensor>(m, "LoDTensor")
+      .def_buffer(
+          [](Tensor &self) -> py::buffer_info { return CastToPyBuffer(self); })
+      .def(
+          "__init__",
+          [](LoDTensor &instance, const std::vector<std::vector<size_t>> &lod) {
+#ifdef PADDLE_ONLY_CPU
+            new (&instance) LoDTensor(lod);
+#else
+             paddle::framework::LoD new_lod;
+             new_lod.reserve(lod.size());
+             std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
+             new (&instance) LoDTensor(new_lod);
+#endif
+          })
+      .def("set_lod",
+           [](LoDTensor &self, const std::vector<std::vector<size_t>> &lod) {
+#ifdef PADDLE_ONLY_CPU
+             self.set_lod(lod);
+#else
+             paddle::framework::LoD new_lod;
+             new_lod.reserve(lod.size());
+             std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
+             self.set_lod(new_lod);
+#endif
+           })
+      .def("lod", [](LoDTensor &self) -> std::vector<std::vector<size_t>> {
+#ifdef PADDLE_ONLY_CPU
+        return self.lod();
+#else
+           auto lod = self.lod();
+           std::vector<std::vector<size_t>> new_lod;
+           new_lod.reserve(lod.size());
+           std::transform(lod.begin(), lod.end(), std::back_inserter(new_lod),
+               [](paddle::framework::Vector<size_t> item) ->
+                   std::vector<size_t> {
+                 std::vector<size_t> v;
+                 v.reserve(item.size());
+                 std::copy(item.begin(), item.end(), std::back_inserter(v));
+                 return v;
+               });
+           return new_lod;
+#endif
+      });
+
   py::class_<Variable>(m, "Variable", R"DOC(Variable Class.
 
 All parameter, weight, gradient are variables in Paddle.
@@ -124,7 +152,9 @@ All parameter, weight, gradient are variables in Paddle.
            [](Variable &var, int val) -> void { *var.GetMutable<int>() = val; })
       .def("get_int", [](const Variable &var) -> int { return var.Get<int>(); })
       .def("get_tensor",
-           [](Variable &self) -> Tensor * { return self.GetMutable<Tensor>(); },
+           [](Variable &self) -> LoDTensor * {
+             return self.GetMutable<LoDTensor>();
+           },
            py::return_value_policy::reference)
       .def("get_net",
            [](Variable &self) -> operators::NetOp * {
@@ -216,7 +246,10 @@ All parameter, weight, gradient are variables in Paddle.
                -> std::map<std::string, std::vector<std::string>> {
                  return op.Outputs();
                })
+      .def("output_vars",
+           [](const OperatorBase &op) { return op.OutputVars(true); })
       .def("inputs", [](const OperatorBase &op) { return op.Inputs(); })
+      .def("input_vars", [](const OperatorBase &op) { return op.InputVars(); })
       .def("__str__", &OperatorBase::DebugString)
       .def("no_intermediate_outputs",
            [](const OperatorBase &op) { return op.OutputVars(false); })
@@ -256,6 +289,28 @@ All parameter, weight, gradient are variables in Paddle.
            [](operators::RecurrentOp &self, const operators::NetOp &net)
                -> void { self.set_stepnet(net.Clone()); });
 
+  // cond_op
+  py::class_<operators::CondOp, OperatorBase>(m, "CondOp")
+      .def_static("create",
+                  [](py::bytes protobin) -> operators::CondOp * {
+                    OpDesc desc;
+                    PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
+                                   "Cannot parse user input to OpDesc");
+                    PADDLE_ENFORCE(desc.IsInitialized(),
+                                   "User OpDesc is not initialized, reason %s",
+                                   desc.InitializationErrorString());
+                    auto cond_op = OpRegistry::CreateOp(desc);
+                    return static_cast<operators::CondOp *>(cond_op.release());
+                  })
+      .def("set_truenet",
+           [](operators::CondOp &self, const operators::NetOp &net) -> void {
+             self.set_truenet(net.Clone());
+           })
+      .def("set_falsenet",
+           [](operators::CondOp &self, const operators::NetOp &net) -> void {
+             self.set_falsenet(net.Clone());
+           });
+
   m.def("unique_integer", UniqueIntegerGenerator);
 
   m.def("is_compile_gpu", IsCompileGPU);
diff --git a/paddle/scripts/docker/build.sh b/paddle/scripts/docker/build.sh
index e57f793ac4..2ac455d771 100644
--- a/paddle/scripts/docker/build.sh
+++ b/paddle/scripts/docker/build.sh
@@ -30,6 +30,8 @@ Configuring cmake in /paddle/build ...
       -DCMAKE_BUILD_TYPE=Release
       -DWITH_DOC=OFF
       -DWITH_GPU=${WITH_GPU:-OFF}
+      -DWITH_MKLDNN=${WITH_MKLDNN:-ON}
+      -DWITH_MKLML=${WITH_MKLML:-ON}
       -DWITH_AVX=${WITH_AVX:-OFF}
       -DWITH_GOLANG=${WITH_GOLANG:-ON}
       -DWITH_SWIG_PY=ON
@@ -50,6 +52,8 @@ cmake .. \
       -DCMAKE_BUILD_TYPE=Release \
       -DWITH_DOC=OFF \
       -DWITH_GPU=${WITH_GPU:-OFF} \
+      -DWITH_MKLDNN=${WITH_MKLDNN:-ON} \
+      -DWITH_MKLML=${WITH_MKLML:-ON} \
       -DWITH_AVX=${WITH_AVX:-OFF} \
       -DWITH_GOLANG=${WITH_GOLANG:-ON} \
       -DWITH_SWIG_PY=${WITH_SWIG_PY:-ON} \
diff --git a/paddle/scripts/docker/build_android.sh b/paddle/scripts/docker/build_android.sh
index aabd2da5e4..11612ad4be 100644
--- a/paddle/scripts/docker/build_android.sh
+++ b/paddle/scripts/docker/build_android.sh
@@ -2,8 +2,30 @@
 
 set -xe
 
+if [ $ANDROID_ABI == "arm64-v8a" ]; then
+  ANDROID_ARCH=arm64
+else # armeabi, armeabi-v7a
+  ANDROID_ARCH=arm
+fi
+
+ANDROID_STANDALONE_TOOLCHAIN=$ANDROID_TOOLCHAINS_DIR/$ANDROID_ARCH-android-$ANDROID_API
+
+cat <<EOF
+============================================
+Generating the standalone toolchain ...
+${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh
+      --arch=$ANDROID_ARCH
+      --platform=android-$ANDROID_API
+      --install-dir=${ANDROID_STANDALONE_TOOLCHAIN}
+============================================
+EOF
+${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh \
+      --arch=$ANDROID_ARCH \
+      --platform=android-$ANDROID_API \
+      --install-dir=$ANDROID_STANDALONE_TOOLCHAIN
+
 BUILD_ROOT=/paddle/build_android
-DEST_ROOT=/paddle/install
+DEST_ROOT=/paddle/install_android
 
 rm -rf $BUILD_ROOT 2>/dev/null || true
 mkdir -p $BUILD_ROOT
@@ -11,7 +33,7 @@ cd $BUILD_ROOT
 
 if [ $ANDROID_ABI == "armeabi-v7a" ]; then
   cmake -DCMAKE_SYSTEM_NAME=Android \
-        -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_ARM_STANDALONE_TOOLCHAIN \
+        -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_STANDALONE_TOOLCHAIN \
         -DANDROID_ABI=$ANDROID_ABI \
         -DANDROID_ARM_NEON=ON \
         -DANDROID_ARM_MODE=ON \
@@ -26,7 +48,7 @@ if [ $ANDROID_ABI == "armeabi-v7a" ]; then
         ..
 elif [ $ANDROID_ABI == "arm64-v8a" ]; then
   cmake -DCMAKE_SYSTEM_NAME=Android \
-        -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_ARM64_STANDALONE_TOOLCHAIN \
+        -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_STANDALONE_TOOLCHAIN \
         -DANDROID_ABI=$ANDROID_ABI \
         -DANDROID_ARM_MODE=ON \
         -DHOST_C_COMPILER=/usr/bin/gcc \
@@ -40,12 +62,12 @@ elif [ $ANDROID_ABI == "arm64-v8a" ]; then
         ..
 elif [ $ANDROID_ABI == "armeabi" ]; then
   cmake -DCMAKE_SYSTEM_NAME=Android \
-        -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_ARM_STANDALONE_TOOLCHAIN \
+        -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_STANDALONE_TOOLCHAIN \
         -DANDROID_ABI=$ANDROID_ABI \
         -DANDROID_ARM_MODE=ON \
         -DHOST_C_COMPILER=/usr/bin/gcc \
         -DHOST_CXX_COMPILER=/usr/bin/g++ \
-        -DCMAKE_INSTALL_PREFIX=/paddle/install \
+        -DCMAKE_INSTALL_PREFIX=$DEST_ROOT \
         -DCMAKE_BUILD_TYPE=Release \
         -DWITH_C_API=ON \
         -DWITH_SWIG_PY=OFF \
@@ -55,5 +77,10 @@ else
   echo "Invalid ANDROID_ABI: $ANDROID_ABI"
 fi
 
+cat <<EOF
+============================================
+Building in $BUILD_ROOT ...
+============================================
+EOF
 make -j `nproc`
 make install -j `nproc`
diff --git a/python/paddle/trainer/config_parser.py b/python/paddle/trainer/config_parser.py
index 7e9112b43b..7c32eb0069 100644
--- a/python/paddle/trainer/config_parser.py
+++ b/python/paddle/trainer/config_parser.py
@@ -2034,6 +2034,7 @@ class ParameterReluLayer(LayerBase):
         config_assert(input_layer.size % partial_sum == 0,
                       "a wrong setting for partial_sum")
         self.set_layer_size(input_layer.size)
+        self.config.partial_sum = partial_sum
         self.create_input_parameter(0, input_layer.size / partial_sum)
 
 
@@ -2054,20 +2055,26 @@ class ConvLayerBase(LayerBase):
         if num_filters is not None:
             self.config.num_filters = num_filters
 
+        use_mkldnn = int(g_command_config_args.get("use_mkldnn", 0))
         use_gpu = int(g_command_config_args.get("use_gpu", 0))
         parallel_nn = int(g_command_config_args.get("parallel_nn", 0))
 
-        # Automatically select cudnn_type for GPU and exconv for CPU
+        # Automatically select cudnn_type for GPU, exconv for CPU
+        # and mkldnn_conv for MKLDNN
         # if set type=conv, but still reserve the way user specify
-        # exconv or cudnn_conv manually.
+        # exconv, mkldnn_conv or cudnn_conv manually.
         if self.layer_type == "cudnn_conv":
             config_assert(use_gpu, "cudnn_conv only support GPU")
 
+        if self.layer_type == "mkldnn_conv":
+            config_assert(use_mkldnn, "mkldnn_conv only support MKLDNN")
+
         if (use_gpu == 1 and self.layer_type != "exconv" and
+                self.layer_type != "mkldnn_conv" and
             (parallel_nn == 0 or self.config.device > -1)):
             self.layer_type = "cudnn_conv"
         else:
-            self.layer_type = "exconv"
+            self.layer_type = "mkldnn_conv" if use_mkldnn else "exconv"
         # need to specify layer in config
         self.config.type = self.layer_type
 
@@ -2099,6 +2106,11 @@ class ConvLayer(ConvLayerBase):
     layer_type = 'exconv'
 
 
+@config_layer('mkldnn_conv')
+class ConvLayer(ConvLayerBase):
+    layer_type = 'mkldnn_conv'
+
+
 @config_layer('cudnn_conv')
 class ConvLayer(ConvLayerBase):
     layer_type = 'cudnn_conv'
@@ -2274,8 +2286,15 @@ class NormLayer(LayerBase):
 
 @config_layer('pool')
 class PoolLayer(LayerBase):
+    layer_type = 'pool'
+
     def __init__(self, name, inputs, ceil_mode=True, **xargs):
-        super(PoolLayer, self).__init__(name, 'pool', 0, inputs=inputs, **xargs)
+        use_mkldnn = int(g_command_config_args.get("use_mkldnn", 0))
+        if self.layer_type == "mkldnn_pool":
+            config_assert(use_mkldnn, "mkldnn_pool only support MKLDNN")
+        self.layer_type = 'mkldnn_pool' if use_mkldnn else 'pool'
+        super(PoolLayer, self).__init__(
+            name, self.layer_type, 0, inputs=inputs, **xargs)
         for input_index in xrange(len(self.inputs)):
             input_layer = self.get_input_layer(input_index)
             pool_conf = self.config.inputs[input_index].pool_conf
@@ -2285,6 +2304,11 @@ class PoolLayer(LayerBase):
                                pool_conf.channels)
 
 
+@config_layer('mkldnn_pool')
+class MKLDNNPoolLayer(PoolLayer):
+    layer_type = 'mkldnn_pool'
+
+
 @config_layer('pool3d')
 class Pool3DLayer(LayerBase):
     def __init__(self, name, inputs, ceil_mode=True, **xargs):
@@ -3748,8 +3772,8 @@ class SwitchOrderLayer(LayerBase):
     def __init__(self, name, inputs, reshape, **xargs):
         super(SwitchOrderLayer, self).__init__(
             name, 'switch_order', 0, inputs=inputs, **xargs)
-        self.config.reshape_conf.heightAxis.extend(reshape['height'])
-        self.config.reshape_conf.widthAxis.extend(reshape['width'])
+        self.config.reshape_conf.height_axis.extend(reshape['height'])
+        self.config.reshape_conf.width_axis.extend(reshape['width'])
 
 
 # Deprecated, use a new layer specific class instead
diff --git a/python/paddle/trainer_config_helpers/layers.py b/python/paddle/trainer_config_helpers/layers.py
index fb1af2d8e8..8c7d1738ad 100644
--- a/python/paddle/trainer_config_helpers/layers.py
+++ b/python/paddle/trainer_config_helpers/layers.py
@@ -169,6 +169,7 @@ class LayerType(object):
     EXCONV_LAYER = 'exconv'
     EXCONVTRANS_LAYER = 'exconvt'
     CUDNNCONV_LAYER = 'cudnn_conv'
+    CUDNNCONVTRANS_LAYER = 'cudnn_convt'
     POOL_LAYER = 'pool'
     POOL3D_LAYER = 'pool3d'
     BATCH_NORM_LAYER = 'batch_norm'
@@ -6461,6 +6462,7 @@ def switch_order_layer(input,
     return LayerOutput(
         name=name,
         layer_type=LayerType.SWITCH_ORDER_LAYER,
+        activation=act,
         parents=input,
         size=l.config.size)
 
diff --git a/python/paddle/trainer_config_helpers/networks.py b/python/paddle/trainer_config_helpers/networks.py
index 34be203ee2..93e8ac173e 100644
--- a/python/paddle/trainer_config_helpers/networks.py
+++ b/python/paddle/trainer_config_helpers/networks.py
@@ -11,10 +11,8 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-"""
 
-"""
-# from activations import *
+
 from activations import LinearActivation, ReluActivation, SoftmaxActivation, \
     IdentityActivation, TanhActivation, SequenceSoftmaxActivation
 from attrs import ExtraAttr
@@ -55,49 +53,49 @@ def sequence_conv_pool(input,
                        context_attr=None,
                        pool_attr=None):
     """
-    Text convolution pooling layers helper.
+    Text convolution pooling group.
 
     Text input => Context Projection => FC Layer => Pooling => Output.
 
-    :param name: name of output layer(pooling layer name)
+    :param name: group name.
     :type name: basestring
-    :param input: name of input layer
+    :param input: input layer.
     :type input: LayerOutput
     :param context_len: context projection length. See
                         context_projection's document.
     :type context_len: int
     :param hidden_size: FC Layer size.
     :type hidden_size: int
-    :param context_start: context projection length. See
+    :param context_start: context start position. See
                           context_projection's context_start.
-    :type context_start: int or None
+    :type context_start: int|None
     :param pool_type: pooling layer type. See pooling_layer's document.
-    :type pool_type: BasePoolingType.
+    :type pool_type: BasePoolingType
     :param context_proj_layer_name: context projection layer name.
                                     None if user don't care.
     :type context_proj_layer_name: basestring
-    :param context_proj_param_attr: context projection parameter attribute.
-                                    None if user don't care.
-    :type context_proj_param_attr: ParameterAttribute or None.
+    :param context_proj_param_attr: padding parameter attribute of context projection layer.
+                                    If false, it means padding always be zero.
+    :type context_proj_param_attr: ParameterAttribute|None
     :param fc_layer_name: fc layer name. None if user don't care.
     :type fc_layer_name: basestring
     :param fc_param_attr: fc layer parameter attribute. None if user don't care.
-    :type fc_param_attr: ParameterAttribute or None
+    :type fc_param_attr: ParameterAttribute|None
     :param fc_bias_attr: fc bias parameter attribute. False if no bias,
                          None if user don't care.
-    :type fc_bias_attr: ParameterAttribute or None
-    :param fc_act: fc layer activation type. None means tanh
+    :type fc_bias_attr: ParameterAttribute|False|None
+    :param fc_act: fc layer activation type. None means tanh.
     :type fc_act: BaseActivation
-    :param pool_bias_attr: pooling layer bias attr. None if don't care.
-                           False if no bias.
-    :type pool_bias_attr: ParameterAttribute or None.
+    :param pool_bias_attr: pooling layer bias attr. False if no bias.
+                           None if user don't care.
+    :type pool_bias_attr: ParameterAttribute|False|None
     :param fc_attr: fc layer extra attribute.
     :type fc_attr: ExtraLayerAttribute
     :param context_attr: context projection layer extra attribute.
     :type context_attr: ExtraLayerAttribute
     :param pool_attr: pooling layer extra attribute.
     :type pool_attr: ExtraLayerAttribute
-    :return: output layer name.
+    :return: layer's output.
     :rtype: LayerOutput
     """
     # Set Default Value to param
@@ -163,45 +161,45 @@ def simple_img_conv_pool(input,
     """
     Simple image convolution and pooling group.
 
-    Input => conv => pooling
+    Img input => Conv => Pooling => Output.
 
-    :param name: group name
+    :param name: group name.
     :type name: basestring
-    :param input: input layer name.
+    :param input: input layer.
     :type input: LayerOutput
-    :param filter_size: see img_conv_layer for details
+    :param filter_size: see img_conv_layer for details.
     :type filter_size: int
-    :param num_filters: see img_conv_layer for details
+    :param num_filters: see img_conv_layer for details.
     :type num_filters: int
-    :param pool_size: see img_pool_layer for details
+    :param pool_size: see img_pool_layer for details.
     :type pool_size: int
-    :param pool_type: see img_pool_layer for details
+    :param pool_type: see img_pool_layer for details.
     :type pool_type: BasePoolingType
-    :param act: see img_conv_layer for details
+    :param act: see img_conv_layer for details.
     :type act: BaseActivation
-    :param groups: see img_conv_layer for details
+    :param groups: see img_conv_layer for details.
     :type groups: int
-    :param conv_stride: see img_conv_layer for details
+    :param conv_stride: see img_conv_layer for details.
     :type conv_stride: int
-    :param conv_padding: see img_conv_layer for details
+    :param conv_padding: see img_conv_layer for details.
     :type conv_padding: int
-    :param bias_attr: see img_conv_layer for details
+    :param bias_attr: see img_conv_layer for details.
     :type bias_attr: ParameterAttribute
-    :param num_channel: see img_conv_layer for details
+    :param num_channel: see img_conv_layer for details.
     :type num_channel: int
-    :param param_attr: see img_conv_layer for details
+    :param param_attr: see img_conv_layer for details.
     :type param_attr: ParameterAttribute
-    :param shared_bias: see img_conv_layer for details
+    :param shared_bias: see img_conv_layer for details.
     :type shared_bias: bool
-    :param conv_layer_attr: see img_conv_layer for details
+    :param conv_layer_attr: see img_conv_layer for details.
     :type conv_layer_attr: ExtraLayerAttribute
-    :param pool_stride: see img_pool_layer for details
+    :param pool_stride: see img_pool_layer for details.
     :type pool_stride: int
-    :param pool_padding: see img_pool_layer for details
+    :param pool_padding: see img_pool_layer for details.
     :type pool_padding: int
-    :param pool_layer_attr: see img_pool_layer for details
+    :param pool_layer_attr: see img_pool_layer for details.
     :type pool_layer_attr: ExtraLayerAttribute
-    :return: Layer's output
+    :return: layer's output
     :rtype: LayerOutput
     """
     _conv_ = img_conv_layer(
@@ -252,48 +250,52 @@ def img_conv_bn_pool(input,
                      pool_layer_attr=None):
     """
     Convolution, batch normalization, pooling group.
+    
+    Img input => Conv => BN => Pooling => Output.
 
-    :param name: group name
+    :param name: group name.
     :type name: basestring
-    :param input: layer's input
-    :type input: LayerOutput
-    :param filter_size: see img_conv_layer's document
+    :param input: input layer.
+    :type input: LayerOutput 
+    :param filter_size: see img_conv_layer for details.
     :type filter_size: int
-    :param num_filters: see img_conv_layer's document
+    :param num_filters: see img_conv_layer for details.
     :type num_filters: int
-    :param pool_size: see img_pool_layer's document.
+    :param pool_size: see img_pool_layer for details.
     :type pool_size: int
-    :param pool_type: see img_pool_layer's document.
+    :param pool_type: see img_pool_layer for details.
     :type pool_type: BasePoolingType
-    :param act: see batch_norm_layer's document.
+    :param act: see batch_norm_layer for details.
     :type act: BaseActivation
-    :param groups: see img_conv_layer's document
+    :param groups: see img_conv_layer for details.
     :type groups: int
-    :param conv_stride: see img_conv_layer's document.
+    :param conv_stride: see img_conv_layer for details.
     :type conv_stride: int
-    :param conv_padding: see img_conv_layer's document.
+    :param conv_padding: see img_conv_layer for details.
     :type conv_padding: int
-    :param conv_bias_attr: see img_conv_layer's document.
+    :param conv_bias_attr: see img_conv_layer for details.
     :type conv_bias_attr: ParameterAttribute
-    :param num_channel: see img_conv_layer's document.
+    :param num_channel: see img_conv_layer for details.
     :type num_channel: int
-    :param conv_param_attr: see img_conv_layer's document.
+    :param conv_param_attr: see img_conv_layer for details.
     :type conv_param_attr: ParameterAttribute
-    :param shared_bias: see img_conv_layer's document.
+    :param shared_bias: see img_conv_layer for details.
     :type shared_bias: bool
-    :param conv_layer_attr: see img_conv_layer's document.
+    :param conv_layer_attr: see img_conv_layer for details.
     :type conv_layer_attr: ExtraLayerOutput
-    :param bn_param_attr: see batch_norm_layer's document.
-    :type bn_param_attr: ParameterAttribute.
-    :param bn_bias_attr: see batch_norm_layer's document.
-    :param bn_layer_attr: ParameterAttribute.
-    :param pool_stride: see img_pool_layer's document.
+    :param bn_param_attr: see batch_norm_layer for details.
+    :type bn_param_attr: ParameterAttribute
+    :param bn_bias_attr: see batch_norm_layer for details.
+    :type bn_bias_attr: ParameterAttribute
+    :param bn_layer_attr: see batch_norm_layer for details.
+    :type bn_layer_attr: ExtraLayerAttribute
+    :param pool_stride: see img_pool_layer for details.
     :type pool_stride: int
-    :param pool_padding: see img_pool_layer's document.
+    :param pool_padding: see img_pool_layer for details.
     :type pool_padding: int
-    :param pool_layer_attr: see img_pool_layer's document.
+    :param pool_layer_attr: see img_pool_layer for details.
     :type pool_layer_attr: ExtraLayerAttribute
-    :return: Layer groups output
+    :return: layer's output
     :rtype: LayerOutput
     """
     __conv__ = img_conv_layer(
@@ -348,10 +350,10 @@ def img_conv_group(input,
     :param conv_batchnorm_drop_rate: if conv_with_batchnorm[i] is true,
         conv_batchnorm_drop_rate[i] represents the drop rate of each batch norm.
     :type conv_batchnorm_drop_rate: list
-    :param input: layer's input.
+    :param input: input layer.
     :type input: LayerOutput
-    :param conv_num_filter: output channels num.
-    :type conv_num_filter: int
+    :param conv_num_filter: list of output channels num.
+    :type conv_num_filter: list|tuple
     :param pool_size: pooling filter size.
     :type pool_size: int
     :param num_channels: input channels num.
@@ -362,18 +364,18 @@ def img_conv_group(input,
     :type conv_filter_size: int
     :param conv_act: activation funciton after convolution.
     :type conv_act: BaseActivation
-    :param conv_with_batchnorm: conv_with_batchnorm[i] represents
-        if there is a batch normalization after each convolution.
+    :param conv_with_batchnorm: if conv_with_batchnorm[i] is true,
+        there is a batch normalization operation after each convolution.
     :type conv_with_batchnorm: list
     :param pool_stride: pooling stride size.
     :type pool_stride: int
     :param pool_type: pooling type.
     :type pool_type: BasePoolingType
-    :param param_attr: Convolution param attribute.
-        None means default attribute.
+    :param param_attr: param attribute of convolution layer,
+                       None means default attribute.
     :type param_attr: ParameterAttribute
-    :return: Layer's output
-    :type: LayerOutput
+    :return: layer's output
+    :rtype: LayerOutput
     """
     tmp = input
 
@@ -466,12 +468,14 @@ def vgg_16_network(input_image, num_channels, num_classes=1000):
     """
     Same model from https://gist.github.com/ksimonyan/211839e770f7b538e2d8
 
-    :param num_classes:
-    :param input_image:
+    :param num_classes: number of class.
+    :type num_classes: int
+    :param input_image: input layer.
     :type input_image: LayerOutput
-    :param num_channels:
+    :param num_channels: input channels num.
     :type num_channels: int
-    :return:
+    :return: layer's output
+    :rtype: LayerOutput
     """
 
     tmp = img_conv_group(
@@ -560,8 +564,8 @@ def simple_lstm(input,
     """
     Simple LSTM Cell.
 
-    It just combine a mixed layer with fully_matrix_projection and a lstmemory
-    layer. The simple lstm cell was implemented as follow equations.
+    It just combines a mixed layer with fully_matrix_projection and a lstmemory
+    layer. The simple lstm cell was implemented with follow equations.
 
     ..  math::
 
@@ -575,37 +579,37 @@ def simple_lstm(input,
 
         h_t & = o_t tanh(c_t)
 
-    Please refer **Generating Sequences With Recurrent Neural Networks** if you
-    want to know what lstm is. Link_ is here.
+    Please refer to **Generating Sequences With Recurrent Neural Networks** for more
+    details about lstm. Link_ is here.
 
     .. _Link: http://arxiv.org/abs/1308.0850
 
     :param name: lstm layer name.
     :type name: basestring
-    :param input: input layer name.
+    :param input: layer's input.
     :type input: LayerOutput
     :param size: lstm layer size.
     :type size: int
-    :param reverse: whether to process the input data in a reverse order
+    :param reverse: process the input in a reverse order or not.
     :type reverse: bool
-    :param mat_param_attr: mixed layer's matrix projection parameter attribute.
+    :param mat_param_attr: parameter attribute of matrix projection in mixed layer.
     :type mat_param_attr: ParameterAttribute
     :param bias_param_attr: bias parameter attribute. False means no bias, None
                             means default bias.
     :type bias_param_attr: ParameterAttribute|False
-    :param inner_param_attr: lstm cell parameter attribute.
+    :param inner_param_attr: parameter attribute of lstm cell.
     :type inner_param_attr: ParameterAttribute
-    :param act: lstm final activiation type
+    :param act: last activiation type of lstm.
     :type act: BaseActivation
-    :param gate_act: lstm gate activiation type
+    :param gate_act: gate activiation type of lstm.
     :type gate_act: BaseActivation
-    :param state_act: lstm state activiation type.
+    :param state_act: state activiation type of lstm.
     :type state_act: BaseActivation
-    :param mixed_layer_attr: mixed layer's extra attribute.
+    :param mixed_layer_attr: extra attribute of mixed layer.
     :type mixed_layer_attr: ExtraLayerAttribute
-    :param lstm_cell_attr: lstm layer's extra attribute.
+    :param lstm_cell_attr: extra attribute of lstm.
     :type lstm_cell_attr: ExtraLayerAttribute
-    :return: lstm layer name.
+    :return: layer's output.
     :rtype: LayerOutput
     """
     fc_name = 'lstm_transform_%s' % name
@@ -643,9 +647,9 @@ def lstmemory_unit(input,
                    lstm_bias_attr=None,
                    lstm_layer_attr=None):
     """
-    Define calculations that a LSTM unit performs during a single time step.
-    This function itself is not a recurrent layer, so it can not be
-    directly used to process sequence inputs. This function is always used in
+    lstmemory_unit defines the caculation process of a LSTM unit during a 
+    single time step. This function is not a recurrent layer, so it can not be
+    directly used to process sequence input. This function is always used in
     recurrent_group (see layers.py for more details) to implement attention
     mechanism.
 
@@ -676,7 +680,7 @@ def lstmemory_unit(input,
                                    state_act=TanhActivation())
 
 
-    :param input: input layer name.
+    :param input: input layer.
     :type input: LayerOutput
     :param out_memory: output of previous time step
     :type out_memory: LayerOutput | None
@@ -684,15 +688,15 @@ def lstmemory_unit(input,
     :type name: basestring
     :param size: lstmemory unit size.
     :type size: int
-    :param param_attr: Parameter config, None if use default.
+    :param param_attr: parameter attribute, None means default attribute.
     :type param_attr: ParameterAttribute
-    :param act: lstm final activiation type
+    :param act: last activiation type of lstm.
     :type act: BaseActivation
-    :param gate_act: lstm gate activiation type
+    :param gate_act: gate activiation type of lstm.
     :type gate_act: BaseActivation
-    :param state_act: lstm state activiation type.
+    :param state_act: state activiation type of lstm.
     :type state_act: BaseActivation
-    :param input_proj_bias_attr: bias attribute for input-to-hidden projection.
+    :param input_proj_bias_attr: bias attribute for input to hidden projection.
                 False means no bias, None means default bias.
     :type input_proj_bias_attr: ParameterAttribute|False|None
     :param input_proj_layer_attr: extra layer attribute for input to hidden
@@ -700,8 +704,8 @@ def lstmemory_unit(input,
     :type input_proj_layer_attr: ExtraLayerAttribute
     :param lstm_bias_attr: bias parameter attribute of lstm layer.
                 False means no bias, None means default bias.
-    :type lstm_bias_attr: ParameterAttribute|False
-    :param lstm_layer_attr: lstm layer's extra attribute.
+    :type lstm_bias_attr: ParameterAttribute|False|None
+    :param lstm_layer_attr: extra attribute of lstm layer.
     :type lstm_layer_attr: ExtraLayerAttribute
     :return: lstmemory unit name.
     :rtype: LayerOutput
@@ -758,9 +762,9 @@ def lstmemory_group(input,
     lstm_group is a recurrent_group version of Long Short Term Memory. It
     does exactly the same calculation as the lstmemory layer (see lstmemory in
     layers.py for the maths) does. A promising benefit is that LSTM memory
-    cell states, or hidden states in every time step are accessible to the
+    cell states(or hidden states) in every time step are accessible to the
     user. This is especially useful in attention model. If you do not need to
-    access the internal states of the lstm, but merely use its outputs,
+    access the internal states of the lstm and merely use its outputs,
     it is recommended to use the lstmemory, which is relatively faster than
     lstmemory_group.
 
@@ -781,28 +785,28 @@ def lstmemory_group(input,
                                     gate_act=SigmoidActivation(),
                                     state_act=TanhActivation())
 
-    :param input: input layer name.
+    :param input: input layer.
     :type input: LayerOutput
     :param size: lstmemory group size.
     :type size: int
-    :param name: name of the lstmemory group.
+    :param name: name of lstmemory group.
     :type name: basestring
-    :param out_memory: output of previous time step
+    :param out_memory: output of previous time step.
     :type out_memory: LayerOutput | None
-    :param reverse: is lstm reversed
+    :param reverse: process the input in a reverse order or not.
     :type reverse: bool
-    :param param_attr: Parameter config, None if use default.
+    :param param_attr: parameter attribute, None means default attribute.
     :type param_attr: ParameterAttribute
-    :param act: lstm final activiation type
+    :param act: last activiation type of lstm.
     :type act: BaseActivation
-    :param gate_act: lstm gate activiation type
+    :param gate_act: gate activiation type of lstm.
     :type gate_act: BaseActivation
-    :param state_act: lstm state activiation type.
+    :param state_act: state activiation type of lstm.
     :type state_act: BaseActivation
     :param lstm_bias_attr: bias parameter attribute of lstm layer.
                            False means no bias, None means default bias.
-    :type lstm_bias_attr: ParameterAttribute|False
-    :param input_proj_bias_attr: bias attribute for input-to-hidden projection.
+    :type lstm_bias_attr: ParameterAttribute|False|None
+    :param input_proj_bias_attr: bias attribute for input to hidden projection.
                 False means no bias, None means default bias.
     :type input_proj_bias_attr: ParameterAttribute|False|None
     :param input_proj_layer_attr: extra layer attribute for input to hidden
@@ -848,15 +852,15 @@ def gru_unit(input,
              gru_layer_attr=None,
              naive=False):
     """
-    Define calculations that a gated recurrent unit performs in a single time
-    step. This function itself is not a recurrent layer, so it can not be
-    directly used to process sequence inputs. This function is always used in
+    gru_unit defines the calculation process of a gated recurrent unit during a single 
+    time step. This function is not a recurrent layer, so it can not be
+    directly used to process sequence input. This function is always used in
     the recurrent_group (see layers.py for more details) to implement attention
     mechanism.
 
     Please see grumemory in layers.py for the details about the maths.
 
-    :param input: input layer name.
+    :param input: input layer.
     :type input: LayerOutput
     :param memory_boot: the initialization state of the LSTM cell.
     :type memory_boot: LayerOutput | None
@@ -864,12 +868,12 @@ def gru_unit(input,
     :type name: basestring
     :param size: hidden size of the gru.
     :type size: int
-    :param act: type of the activation
+    :param act: activation type of gru
     :type act: BaseActivation
-    :param gate_act: type of the gate activation
+    :param gate_act: gate activation type or gru
     :type gate_act: BaseActivation
-    :param gru_layer_attr: Extra parameter attribute of the gru layer.
-    :type gru_layer_attr: ParameterAttribute|False
+    :param gru_layer_attr: Extra attribute of the gru layer.
+    :type gru_layer_attr: ExtraLayerAttribute
     :return: the gru output layer.
     :rtype: LayerOutput
     """
@@ -915,7 +919,7 @@ def gru_group(input,
     does exactly the same calculation as the grumemory layer does. A promising
     benefit is that gru hidden states are accessible to the user. This is
     especially useful in attention model. If you do not need to access
-    any internal state, but merely use the outputs of a GRU, it is recommended
+    any internal state and merely use the outputs of a GRU, it is recommended
     to use the grumemory, which is relatively faster.
 
     Please see grumemory in layers.py for more detail about the maths.
@@ -924,12 +928,12 @@ def gru_group(input,
 
     ..  code-block:: python
 
-        gru = gur_group(input=[layer1],
+        gru = gru_group(input=[layer1],
                         size=256,
                         act=TanhActivation(),
                         gate_act=SigmoidActivation())
 
-    :param input: input layer name.
+    :param input: input layer.
     :type input: LayerOutput
     :param memory_boot: the initialization state of the LSTM cell.
     :type memory_boot: LayerOutput | None
@@ -937,16 +941,17 @@ def gru_group(input,
     :type name: basestring
     :param size: hidden size of the gru.
     :type size: int
-    :param reverse: whether to process the input data in a reverse order
+    :param reverse: process the input in a reverse order or not.
     :type reverse: bool
-    :param act: type of the activiation
+    :param act: activiation type of gru
     :type act: BaseActivation
-    :param gate_act: type of the gate activiation
+    :param gate_act: gate activiation type of gru
     :type gate_act: BaseActivation
-    :param gru_bias_attr: bias. False means no bias, None means default bias.
-    :type gru_bias_attr: ParameterAttribute|False
-    :param gru_layer_attr: Extra parameter attribute of the gru layer.
-    :type gru_layer_attr: ParameterAttribute|False
+    :param gru_bias_attr: bias parameter attribute of gru layer,
+                          False means no bias, None means default bias.
+    :type gru_bias_attr: ParameterAttribute|False|None
+    :param gru_layer_attr: Extra attribute of the gru layer.
+    :type gru_layer_attr: ExtraLayerAttribute
     :return: the gru group.
     :rtype: LayerOutput
     """
@@ -986,11 +991,11 @@ def simple_gru(input,
                gru_layer_attr=None,
                naive=False):
     """
-    You maybe see gru_step_layer, grumemory in layers.py, gru_unit, gru_group,
+    You may see gru_step_layer, grumemory in layers.py, gru_unit, gru_group,
     simple_gru in network.py. The reason why there are so many interfaces is
     that we have two ways to implement recurrent neural network. One way is to
     use one complete layer to implement rnn (including simple rnn, gru and lstm)
-    with multiple time steps, such as recurrent_layer, lstmemory, grumemory. But,
+    with multiple time steps, such as recurrent_layer, lstmemory, grumemory. But 
     the multiplication operation :math:`W x_t` is not computed in these layers.
     See details in their interfaces in layers.py.
     The other implementation is to use an recurrent group which can ensemble a
@@ -1018,22 +1023,23 @@ def simple_gru(input,
 
         gru = simple_gru(input=[layer1], size=256)
 
-    :param input: input layer name.
+    :param input: input layer.
     :type input: LayerOutput
     :param name: name of the gru group.
     :type name: basestring
     :param size: hidden size of the gru.
     :type size: int
-    :param reverse: whether to process the input data in a reverse order
+    :param reverse: process the input in a reverse order or not.
     :type reverse: bool
-    :param act: type of the activiation
+    :param act: activiation type of gru
     :type act: BaseActivation
-    :param gate_act: type of the gate activiation
+    :param gate_act: gate activiation type of gru
     :type gate_act: BaseActivation
-    :param gru_bias_attr: bias. False means no bias, None means default bias.
-    :type gru_bias_attr: ParameterAttribute|False
-    :param gru_layer_attr: Extra parameter attribute of the gru layer.
-    :type gru_layer_attr: ParameterAttribute|False
+    :param gru_bias_attr: bias parameter attribute of gru layer,
+                          False means no bias, None means default bias.
+    :type gru_bias_attr: ParameterAttribute|False|None
+    :param gru_layer_attr: Extra attribute of the gru layer.
+    :type gru_layer_attr: ExtraLayerAttribute
     :return: the gru group.
     :rtype: LayerOutput
     """
@@ -1071,8 +1077,8 @@ def simple_gru2(input,
                 mixed_layer_attr=None,
                 gru_cell_attr=None):
     """
-    simple_gru2 is the same with simple_gru, but using grumemory instead
-    Please see grumemory in layers.py for more detail about the maths.
+    simple_gru2 is the same with simple_gru, but using grumemory instead.
+    Please refer to grumemory in layers.py for more detail about the math.
     simple_gru2 is faster than simple_gru.
 
     The example usage is:
@@ -1081,22 +1087,23 @@ def simple_gru2(input,
 
         gru = simple_gru2(input=[layer1], size=256)
 
-    :param input: input layer name.
+    :param input: input layer.
     :type input: LayerOutput
     :param name: name of the gru group.
     :type name: basestring
     :param size: hidden size of the gru.
     :type size: int
-    :param reverse: whether to process the input data in a reverse order
+    :param reverse: process the input in a reverse order or not.
     :type reverse: bool
-    :param act: type of the activiation
+    :param act: activiation type of gru
     :type act: BaseActivation
-    :param gate_act: type of the gate activiation
+    :param gate_act: gate activiation type of gru
     :type gate_act: BaseActivation
-    :param gru_bias_attr: bias. False means no bias, None means default bias.
-    :type gru_bias_attr: ParameterAttribute|False
-    :param gru_layer_attr: Extra parameter attribute of the gru layer.
-    :type gru_layer_attr: ParameterAttribute|False
+    :param gru_bias_attr: bias parameter attribute of gru layer, 
+                          False means no bias, None means default bias.
+    :type gru_bias_attr: ParameterAttribute|False|None
+    :param gru_layer_attr: Extra attribute of the gru layer.
+    :type gru_layer_attr: ExtraLayerAttribute
     :return: the gru group.
     :rtype: LayerOutput
     """
@@ -1145,7 +1152,7 @@ def bidirectional_gru(input,
                       concat_act=None):
     """
     A bidirectional_gru is a recurrent unit that iterates over the input
-    sequence both in forward and bardward orders, and then concatenate two
+    sequence both in forward and backward orders, and then concatenate two
     outputs to form a final output. However, concatenation of two outputs
     is not the only way to form the final output, you can also, for example,
     just add them together.
@@ -1162,11 +1169,10 @@ def bidirectional_gru(input,
     :type input: LayerOutput
     :param size: gru layer size.
     :type size: int
-    :param return_seq: If set False, outputs of the last time step are
-                       concatenated and returned.
-                       If set True, the entire output sequences that are
-                       processed in forward and backward directions are
+    :param return_seq: If set False, the last time step of output are
                        concatenated and returned.
+                       If set True, the entire output sequences in forward 
+                       and backward directions are concatenated and returned.
     :type return_seq: bool
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -1230,8 +1236,8 @@ def bidirectional_lstm(input,
                        concat_act=None):
     """
     A bidirectional_lstm is a recurrent unit that iterates over the input
-    sequence both in forward and bardward orders, and then concatenate two
-    outputs form a final output. However, concatenation of two outputs
+    sequence both in forward and backward orders, and then concatenate two
+    outputs to form a final output. However, concatenation of two outputs
     is not the only way to form the final output, you can also, for example,
     just add them together.
 
@@ -1252,13 +1258,12 @@ def bidirectional_lstm(input,
     :type input: LayerOutput
     :param size: lstm layer size.
     :type size: int
-    :param return_seq: If set False, outputs of the last time step are
-                       concatenated and returned.
-                       If set True, the entire output sequences that are
-                       processed in forward and backward directions are
+    :param return_seq: If set False, the last time step of output are
                        concatenated and returned.
+                       If set True, the entire output sequences in forward 
+                       and backward directions are concatenated and returned.
     :type return_seq: bool
-    :return: LayerOutput object accroding to the return_seq.
+    :return: LayerOutput object.
     :rtype: LayerOutput
     """
     args = locals()
@@ -1303,7 +1308,7 @@ def simple_attention(encoded_sequence,
                      weight_act=None,
                      name=None):
     """
-    Calculate and then return a context vector by attention machanism.
+    Calculate and return a context vector with attention mechanism.
     Size of the context vector equals to size of the encoded_sequence.
 
     ..  math::
@@ -1336,10 +1341,10 @@ def simple_attention(encoded_sequence,
     :param name: name of the attention model.
     :type name: basestring
     :param softmax_param_attr: parameter attribute of sequence softmax
-                               that is used to produce attention weight
+                               that is used to produce attention weight.
     :type softmax_param_attr: ParameterAttribute
-    :param weight_act: activation of the attention model
-    :type weight_act: Activation
+    :param weight_act: activation of the attention model.
+    :type weight_act: BaseActivation
     :param encoded_sequence: output of the encoder
     :type encoded_sequence: LayerOutput
     :param encoded_proj: attention weight is computed by a feed forward neural
@@ -1411,7 +1416,7 @@ def inputs(layers, *args):
 
 def outputs(layers, *args):
     """
-    Declare the outputs of network. If user have not defined the inputs of
+    Declare the outputs of network. If user has not defined the inputs of
     network, this method will calculate the input order by dfs travel.
 
     :param layers: Output layers.
diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_prelu_layer.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_prelu_layer.protostr
index 64d227565f..94ad56cab0 100644
--- a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_prelu_layer.protostr
+++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_prelu_layer.protostr
@@ -14,6 +14,29 @@ layers {
     input_layer_name: "input"
     input_parameter_name: "___prelu_layer_0__.w0"
   }
+  partial_sum: 1
+}
+layers {
+  name: "__prelu_layer_1__"
+  type: "prelu"
+  size: 300
+  active_type: ""
+  inputs {
+    input_layer_name: "input"
+    input_parameter_name: "___prelu_layer_1__.w0"
+  }
+  partial_sum: 1
+}
+layers {
+  name: "__prelu_layer_2__"
+  type: "prelu"
+  size: 300
+  active_type: ""
+  inputs {
+    input_layer_name: "input"
+    input_parameter_name: "___prelu_layer_2__.w0"
+  }
+  partial_sum: 5
 }
 parameters {
   name: "___prelu_layer_0__.w0"
@@ -23,14 +46,32 @@ parameters {
   initial_strategy: 0
   initial_smart: true
 }
+parameters {
+  name: "___prelu_layer_1__.w0"
+  size: 300
+  initial_mean: 0.0
+  initial_std: 0.057735026919
+  initial_strategy: 0
+  initial_smart: true
+}
+parameters {
+  name: "___prelu_layer_2__.w0"
+  size: 60
+  initial_mean: 0.0
+  initial_std: 0.129099444874
+  initial_strategy: 0
+  initial_smart: true
+}
 input_layer_names: "input"
-output_layer_names: "__prelu_layer_0__"
+output_layer_names: "__prelu_layer_2__"
 sub_models {
   name: "root"
   layer_names: "input"
   layer_names: "__prelu_layer_0__"
+  layer_names: "__prelu_layer_1__"
+  layer_names: "__prelu_layer_2__"
   input_layer_names: "input"
-  output_layer_names: "__prelu_layer_0__"
+  output_layer_names: "__prelu_layer_2__"
   is_recurrent_layer_group: false
 }
 
diff --git a/python/paddle/trainer_config_helpers/tests/configs/test_prelu_layer.py b/python/paddle/trainer_config_helpers/tests/configs/test_prelu_layer.py
index 2e3057f323..aae90fab32 100644
--- a/python/paddle/trainer_config_helpers/tests/configs/test_prelu_layer.py
+++ b/python/paddle/trainer_config_helpers/tests/configs/test_prelu_layer.py
@@ -2,5 +2,7 @@ from paddle.trainer_config_helpers import *
 
 data = data_layer(name='input', size=300)
 prelu = prelu_layer(input=data)
+prelu = prelu_layer(input=data, partial_sum=1)
+prelu = prelu_layer(input=data, partial_sum=5)
 
 outputs(prelu)
diff --git a/python/paddle/v2/event.py b/python/paddle/v2/event.py
index 7589cc9917..e66bf67d79 100644
--- a/python/paddle/v2/event.py
+++ b/python/paddle/v2/event.py
@@ -53,10 +53,13 @@ class BeginPass(object):
 class EndPass(WithMetric):
     """
     Event On One Pass Training Complete.
+    To get the output of a specific layer, add "event.gm.getLayerOutputs('predict_layer')"
+    in your event_handler call back
     """
 
-    def __init__(self, pass_id, evaluator):
+    def __init__(self, pass_id, evaluator, gm):
         self.pass_id = pass_id
+        self.gm = gm
         WithMetric.__init__(self, evaluator)
 
 
@@ -73,10 +76,13 @@ class BeginIteration(object):
 class EndIteration(WithMetric):
     """
     Event On One Batch Training Complete.
+    To get the output of a specific layer, add "event.gm.getLayerOutputs('predict_layer')"
+    in your event_handler call back
     """
 
-    def __init__(self, pass_id, batch_id, cost, evaluator):
+    def __init__(self, pass_id, batch_id, cost, evaluator, gm):
         self.pass_id = pass_id
         self.batch_id = batch_id
         self.cost = cost
+        self.gm = gm
         WithMetric.__init__(self, evaluator)
diff --git a/python/paddle/v2/framework/op.py b/python/paddle/v2/framework/op.py
index c1585bcffc..6cca41e43b 100644
--- a/python/paddle/v2/framework/op.py
+++ b/python/paddle/v2/framework/op.py
@@ -43,7 +43,6 @@ class OpDescCreationMethod(object):
         if len(args) != 0:
             raise ValueError("Only keyword arguments are supported.")
         op_desc = framework_pb2.OpDesc()
-
         for input_parameter in self.__op_proto__.inputs:
             input_arguments = kwargs.get(input_parameter.name, [])
             if is_str(input_arguments):
@@ -98,7 +97,7 @@ class OpDescCreationMethod(object):
                     new_attr.strings.extend(user_defined_attr)
                 elif attr.type == framework_pb2.INT_PAIRS:
                     for p in user_defined_attr:
-                        pair = new_attr.pairs.add()
+                        pair = new_attr.int_pairs.add()
                         pair.first = p[0]
                         pair.second = p[1]
                 else:
@@ -142,8 +141,8 @@ def create_op_creation_method(op_proto):
     return OpInfo(
         method=__impl__,
         name=op_proto.type,
-        inputs=[var.name for var in op_proto.inputs],
-        outputs=[var.name for var in op_proto.outputs],
+        inputs=[(var.name, var.duplicable) for var in op_proto.inputs],
+        outputs=[(var.name, var.duplicable) for var in op_proto.outputs],
         attrs=[attr.name for attr in op_proto.attrs])
 
 
@@ -180,9 +179,15 @@ class OperatorFactory(object):
         return self.op_methods.get(t)
 
     def get_op_input_names(self, type):
+        return map(lambda x: x[0], self.get_op_info(type).inputs)
+
+    def get_op_inputs(self, type):
         return self.get_op_info(type).inputs
 
     def get_op_output_names(self, type):
+        return map(lambda x: x[0], self.get_op_info(type).outputs)
+
+    def get_op_outputs(self, type):
         return self.get_op_info(type).outputs
 
     def get_op_attr_names(self, type):
@@ -210,5 +215,27 @@ class __RecurrentOp__(object):
         return core.RecurrentOp.create(proto.SerializeToString())
 
 
+class __CondOp__(object):
+    __proto__ = None
+    type = "cond"
+
+    def __init__(self):
+        # cache recurrent_op's proto
+        if self.__proto__ is None:
+            for op_proto in get_all_op_protos():
+                if op_proto.type == self.type:
+                    self.__proto__ = op_proto
+
+    def __call__(self, *args, **kwargs):
+        if self.type not in args and "type" not in kwargs:
+            kwargs["type"] = self.type
+        # create proto
+        create_method = OpDescCreationMethod(self.__proto__)
+        proto = create_method(*args, **kwargs)
+        # create condop
+        return core.CondOp.create(proto.SerializeToString())
+
+
 Operator = OperatorFactory()  # The default global factory
 RecurrentOp = __RecurrentOp__()
+CondOp = __CondOp__()
diff --git a/python/paddle/v2/framework/tests/CMakeLists.txt b/python/paddle/v2/framework/tests/CMakeLists.txt
index ef910f939b..4d7664469e 100644
--- a/python/paddle/v2/framework/tests/CMakeLists.txt
+++ b/python/paddle/v2/framework/tests/CMakeLists.txt
@@ -1,37 +1,5 @@
-py_test(test_net SRCS test_net.py)
-
-py_test(test_scope SRCS test_scope.py)
-
-py_test(test_tensor SRCS test_tensor.py)
-py_test(test_mul_op SRCS test_mul_op.py)
-py_test(test_cos_sim_op SRCS test_cos_sim_op.py)
-
-py_test(test_mean_op SRCS test_mean_op.py)
-
-py_test(test_protobuf SRCS test_protobuf.py)
-
-py_test(test_add_two_op SRCS test_add_two_op.py)
-py_test(test_sigmoid_op SRCS test_sigmoid_op.py)
-py_test(test_softmax_op SRCS test_softmax_op.py)
-py_test(test_cross_entropy_op SRCS test_cross_entropy_op.py)
-py_test(test_gather_op SRCS test_gather_op.py)
-py_test(test_scatter_op SRCS test_scatter_op.py)
-py_test(test_fill_zeros_like_op SRCS test_fill_zeros_like_op.py)
-py_test(test_top_k_op SRCS test_top_k_op.py)
-
-py_test(gradient_checker SRCS gradient_checker.py)
-
-py_test(test_rowwise_add_op SRCS test_rowwise_add_op.py)
-
-py_test(test_default_scope_funcs SRCS test_default_scope_funcs.py)
-
-py_test(test_operator SRCS test_operator.py)
-py_test(test_gaussian_random_op SRCS test_gaussian_random_op.py)
-py_test(test_uniform_random_op SRCS test_uniform_random_op.py)
-py_test(test_recurrent_op SRCS test_recurrent_op.py)
-py_test(test_sgd_op SRCS test_sgd_op.py)
-py_test(test_gradient_checker SRCS test_gradient_checker.py)
-py_test(test_lookup_table SRCS test_lookup_table.py)
-py_test(test_scale_and_identity_op SRCS test_scale_and_identity_op.py)
-py_test(mnist SRCS mnist.py)
-py_test(test_squared_l2_distance_op SRCS test_squared_l2_distance_op.py)
+file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
+string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
+foreach(src ${TEST_OPS})
+    py_test(${src} SRCS ${src}.py)
+endforeach()
diff --git a/python/paddle/v2/framework/tests/gradient_checker.py b/python/paddle/v2/framework/tests/gradient_checker.py
deleted file mode 100644
index fdb06b7988..0000000000
--- a/python/paddle/v2/framework/tests/gradient_checker.py
+++ /dev/null
@@ -1,311 +0,0 @@
-import unittest
-
-import numpy
-import itertools
-import paddle.v2.framework.core as core
-from paddle.v2.framework.op import Operator
-
-__all__ = ['get_numeric_gradient']
-
-
-def create_op(op_type):
-    # TODO need to set attrs
-    kwargs = dict()
-    for in_name in Operator.get_op_input_names(op_type):
-        kwargs[in_name] = in_name
-    for out_name in Operator.get_op_output_names(op_type):
-        kwargs[out_name] = out_name
-
-    return Operator(op_type, **kwargs)
-
-
-def grad_var_name(var_name):
-    return var_name + "@GRAD"
-
-
-def empty_var_name():
-    return "@EMPTY@"
-
-
-def get_numeric_gradient(op,
-                         input_values,
-                         output_name,
-                         input_to_check,
-                         delta=0.005,
-                         local_scope=None,
-                         in_place=False):
-    """
-    Get Numeric Gradient for an operator's input.
-
-    :param op: C++ operator instance, could be an network
-    :param input_values: The input variables. Should be an dictionary, key is
-    variable name. Value is numpy array.
-    :param output_name: The final output variable name.
-    :param input_to_check: The input variable need to get gradient.
-    :param delta: The perturbation value for numeric gradient method. The
-    smaller delta is, the more accurate result will get. But if that delta is
-     too small, it could occur numerical stability problem.
-    :param local_scope: The local scope used for get_numeric_gradient.
-    :return: The gradient array in numpy format.
-    """
-    if local_scope is None:
-        local_scope = core.Scope()
-
-    # Create all input variable in local_scope
-    for var_name in input_values:
-        var = local_scope.new_var(var_name)
-        tensor = var.get_tensor()
-        tensor.set_dims(input_values[var_name].shape)
-        tensor.alloc_float(core.CPUPlace())
-        tensor.set(input_values[var_name], core.CPUPlace())
-
-    # Create all output variable in local_scope
-    opts = op.outputs()
-    for key in opts:
-        for output in opts[key]:
-            if local_scope.find_var(output) is None:
-                local_scope.new_var(output).get_tensor()
-    op.infer_shape(local_scope)
-
-    # allocate output memory
-    for key in opts:
-        for output in opts[key]:
-            local_scope.find_var(output).get_tensor().alloc_float(core.CPUPlace(
-            ))
-
-    cpu_ctx = core.DeviceContext.create(core.CPUPlace())
-
-    def get_output():
-        op.run(local_scope, cpu_ctx)
-        return numpy.array(local_scope.find_var(output_name).get_tensor()).sum()
-
-    def product(dim):
-        return reduce(lambda a, b: a * b, dim, 1)
-
-    def restore_inputs():
-        for var_name in input_values:
-            tensor_ = local_scope.find_var(var_name).get_tensor()
-            tensor_.set(numpy.copy(input_values[var_name]), core.CPUPlace())
-
-    # get the input tensor that we want to get it's numeric gradient.
-    tensor_to_check = local_scope.find_var(input_to_check).get_tensor()
-    tensor_size = product(tensor_to_check.get_dims())
-    # prepare a numpy array to store the gradient.
-    gradient_flat = numpy.zeros(shape=(tensor_size, ), dtype='float32')
-
-    # we only compute gradient of one element each time.
-    # we use a for loop to compute the gradient of every element.
-    for i in xrange(tensor_size):
-        if in_place:
-            restore_inputs()
-        # get one input element throw it's index i.
-        origin = tensor_to_check.get_float_element(i)
-
-        # add delta to it, run op and then get the sum of the result tensor.
-        x_pos = origin + delta
-        tensor_to_check.set_float_element(i, x_pos)
-        y_pos = get_output()
-
-        # plus delta to this element, run op and get the sum of the result tensor.
-        if in_place:
-            restore_inputs()
-        x_neg = origin - delta
-        tensor_to_check.set_float_element(i, x_neg)
-        y_neg = get_output()
-
-        # restore old value
-        tensor_to_check.set_float_element(i, origin)
-
-        # compute the gradient of this element and store it into a numpy array.
-        gradient_flat[i] = (y_pos - y_neg) / delta / 2
-
-    # reshape the gradient result to the shape of the source tensor.
-    return gradient_flat.reshape(tensor_to_check.get_dims())
-
-
-class GradientChecker(unittest.TestCase):
-    def __get_gradient(self, forward_op, backward_op, input_value, grad_names,
-                       place):
-        """Get the input gradients after running forward and backward operators
-        on the given places.
-
-        :param forward_op: forward operator
-        :type forward_op: Operator
-        :param backward_op: backward operator
-        :type backward_op: Operator
-        :param input_value: input values.
-        :type input_value: dict{string:numpy.array}
-        :param grad_names: the names of returned input gradients.
-        :type input_value: a list of string
-        :param place: the device type.
-        :type place: CPUPlace or GPUPlace
-        :return: the input grdients of given grad_names.
-        :rtype: a list of numpy.array
-        """
-        scope = core.Scope()
-        ctx = core.DeviceContext.create(place)
-
-        inputs = forward_op.inputs()
-        in_names = [item for k in inputs for item in inputs[k]]
-        outputs = forward_op.outputs()
-        out_names = [item for k in outputs for item in outputs[k]]
-
-        # create input var and set value
-        for name, value in input_value.iteritems():
-            if name not in in_names:
-                raise ValueError(name + "does not exist in Op's inputs.")
-            var = scope.new_var(name).get_tensor()
-            var.set_dims(value.shape)
-            var.set(value, place)
-
-        # run forward op
-        for out_name in out_names:
-            scope.new_var(out_name)
-        forward_op.infer_shape(scope)
-        forward_op.run(scope, ctx)
-
-        # set output var's shape
-        # set output grad to ones
-        for name in out_names:
-            out_tensor = scope.find_var(name).get_tensor()
-            grad_tensor = scope.new_var(grad_var_name(name)).get_tensor()
-            grad_tensor.set_dims(out_tensor.shape())
-            data = numpy.ones(out_tensor.shape(), dtype=numpy.float32)
-            grad_tensor.set(data, place)
-
-        # run backward op
-        backward_outs = backward_op.outputs()
-        backward_names = [
-            item for key in backward_outs for item in backward_outs[key]
-        ]
-        for name in backward_names:
-            scope.new_var(name)
-
-        backward_op.infer_shape(scope)
-        backward_op.run(scope, ctx)
-
-        outs = [
-            numpy.array(scope.find_var(name).get_tensor())
-            for name in grad_names
-        ]
-        return outs
-
-    def compare_grad(self, forward_op, input_value, no_grad_set=None):
-        """ Compare the input gradients between CPU and GPU for the given forward
-        operator.
-
-        :param forward_op: forward operator
-        :type forward_op: Operator
-        :param input_value: input values.
-        :type input_value: dict{string:numpy.array}
-        :param no_grad_set: the set of variables names without gradients.
-        :type no_grad_set: a set of string
-        :raises: AssertionError, there is different gradient value.
-        """
-        if no_grad_set is None:
-            no_grad_set = set()
-        backward_op = core.Operator.backward(forward_op, no_grad_set)
-        # return if not compile with GPU or not implementing GPU kernel
-        if not (core.is_compile_gpu() and backward_op.support_gpu()):
-            return
-
-        outputs = backward_op.outputs()
-        out_names = [item for k in outputs for item in outputs[k]]
-        out_names = filter(lambda x: x != empty_var_name(), out_names)
-        cpu_grads = self.__get_gradient(forward_op, backward_op, input_value,
-                                        out_names, core.CPUPlace())
-        gpu_grads = self.__get_gradient(forward_op, backward_op, input_value,
-                                        out_names, core.GPUPlace(0))
-
-        for c_grad, g_grad, name in itertools.izip(cpu_grads, gpu_grads,
-                                                   out_names):
-            self.assertTrue(
-                numpy.allclose(
-                    c_grad, g_grad, atol=1e-4),
-                "output name: " + name + " has diff")
-
-    def __assert_is_close(self, numeric_grads, analytic_grads, names,
-                          max_relative_error, msg_prefix):
-        """Use relative error for the comparison.
-
-        :param numeric_grads: the numerical graidents.
-        :type numeric_grads: a list of numpy.array
-        :param analytic_grads: the analytical graidents.
-        :type analytic_grads: a list of numpy.array
-        :param name: the names of gradients, used to print for debug.
-        :type names: a list of string
-        :param msg_prefix: string info, used to print for debug.
-        :type msf_prefix: string
-        """
-        for a, b, name in itertools.izip(numeric_grads, analytic_grads, names):
-            abs_a = numpy.abs(a)
-            # if abs_a is nearly zero, then use abs error for a, not relative
-            # error.
-            abs_a[abs_a < 1e-3] = 1
-
-            diff_mat = numpy.abs(a - b) / abs_a
-            max_diff = numpy.max(diff_mat)
-
-            def err_msg():
-                offset = numpy.argmax(diff_mat > max_relative_error)
-                return "%s Variable %s max gradient diff %f over limit %f, the first " \
-                       "error element is %d" % (
-                       msg_prefix, name, max_diff, max_relative_error, offset)
-
-            self.assertLessEqual(max_diff, max_relative_error, err_msg())
-
-    def check_grad(self,
-                   forward_op,
-                   input_vars,
-                   inputs_to_check,
-                   output_name,
-                   no_grad_set=None,
-                   only_cpu=False,
-                   in_place=False,
-                   max_relative_error=0.005):
-        """
-        :param forward_op: used to create backward_op
-        :param input_vars: numpy value of input variable. The following
-            computation will use these variables.
-        :param inputs_to_check: inputs var names that should check gradient.
-        :param output_name: the output variable name of forward network.
-        :param max_relative_error: The relative tolerance parameter.
-        :param no_grad_set: used when create backward ops
-        :param only_cpu: only compute and check gradient on cpu kernel.
-        :return:
-        """
-        if no_grad_set is None:
-            no_grad_set = set()
-
-        no_tmp_out = forward_op.no_intermediate_outputs()
-        if len(no_tmp_out) != 1:
-            raise ValueError("non temp out_names should be 1")
-
-        inputs = forward_op.inputs()
-        in_names = [item for k in inputs for item in inputs[k]]
-        for no_grad in no_grad_set:
-            if no_grad not in in_names:
-                raise ValueError("no_grad should be in in_names")
-            if no_grad in inputs_to_check:
-                raise ValueError("no_grad should not be in inputs_to_check")
-
-        backward_op = core.Operator.backward(forward_op, no_grad_set)
-
-        places = [core.CPUPlace()]
-        if not only_cpu and core.is_compile_gpu() and backward_op.support_gpu():
-            places.append(core.GPUPlace(0))
-
-        # get numerical gradients
-        numeric_grads = [
-            get_numeric_gradient(
-                forward_op, input_vars, output_name, name, in_place=in_place)
-            for name in inputs_to_check
-        ]
-
-        check_names = [grad_var_name(name) for name in inputs_to_check]
-        for place in places:
-            analytic_grads = self.__get_gradient(forward_op, backward_op,
-                                                 input_vars, check_names, place)
-            self.__assert_is_close(numeric_grads, analytic_grads, check_names,
-                                   max_relative_error,
-                                   "Gradient Check On %s" % str(place))
diff --git a/python/paddle/v2/framework/tests/op_test.py b/python/paddle/v2/framework/tests/op_test.py
new file mode 100644
index 0000000000..0a5673868c
--- /dev/null
+++ b/python/paddle/v2/framework/tests/op_test.py
@@ -0,0 +1,303 @@
+import unittest
+import numpy as np
+import itertools
+import paddle.v2.framework.core as core
+from paddle.v2.framework.op import Operator
+
+
+def grad_var_name(var_name):
+    return var_name + "@GRAD"
+
+
+def create_op(scope, op_type, inputs, outputs, attrs):
+    kwargs = dict()
+
+    for in_name, in_dup in Operator.get_op_inputs(op_type):
+        if in_name in inputs:
+            kwargs[in_name] = []
+            if in_dup:
+                sub_in = inputs[in_name]
+                for sub_in_name, _ in sub_in:
+                    var = scope.new_var(sub_in_name)
+                    kwargs[in_name].append(sub_in_name)
+            else:
+                var = scope.new_var(in_name)
+                kwargs[in_name].append(in_name)
+
+    for out_name, out_dup in Operator.get_op_outputs(op_type):
+        if out_name in outputs:
+            kwargs[out_name] = []
+            if out_dup:
+                sub_out = outputs[out_name]
+                for sub_out_name, _ in sub_out:
+                    var = scope.new_var(sub_out_name)
+                    kwargs[out_name].append(sub_out_name)
+            else:
+                var = scope.new_var(out_name)
+                kwargs[out_name].append(out_name)
+
+    for attr_name in Operator.get_op_attr_names(op_type):
+        if attr_name in attrs:
+            kwargs[attr_name] = attrs[attr_name]
+
+    return Operator(op_type, **kwargs)
+
+
+def set_input(scope, op, inputs, place):
+    for in_name, in_dup in Operator.get_op_inputs(op.type()):
+        if in_name in inputs:
+            if in_dup:
+                sub_in = inputs[in_name]
+                for sub_in_name, sub_in_val in sub_in:
+                    var = scope.find_var(sub_in_name)
+                    tensor = var.get_tensor()
+                    sub_in_array = sub_in_val[0] \
+                        if isinstance(sub_in_val, tuple) else sub_in_val
+                    tensor.set_dims(sub_in_array.shape)
+                    tensor.set(sub_in_array, place)
+                    if isinstance(sub_in_val, tuple):
+                        tensor.set_lod(sub_in_val[1])
+            else:
+                var = scope.find_var(in_name)
+                tensor = var.get_tensor()
+                in_val = inputs[in_name]
+                in_array = in_val[0] if isinstance(in_val, tuple) else in_val
+                tensor.set_dims(in_array.shape)
+                tensor.set(in_array, place)
+                if isinstance(in_val, tuple):
+                    tensor.set_lod(in_val[1])
+
+
+def set_output_grad(scope, op, outputs, place):
+    for out_name, out_dup in Operator.get_op_outputs(op.type()):
+        if out_name in outputs:
+            if out_dup:
+                sub_out = outputs[out_name]
+                for sub_out_name, _ in sub_out:
+                    out_tensor = scope.find_var(sub_out_name).get_tensor()
+                    grad_tensor = scope.new_var(grad_var_name(
+                        sub_out_name)).get_tensor()
+                    grad_tensor.set_dims(out_tensor.shape())
+                    data = np.ones(out_tensor.shape(), dtype=np.float32)
+                    grad_tensor.set(data, place)
+            else:
+                out_tensor = scope.find_var(out_name).get_tensor()
+                grad_tensor = scope.new_var(grad_var_name(out_name)).get_tensor(
+                )
+                grad_tensor.set_dims(out_tensor.shape())
+                data = np.ones(out_tensor.shape(), dtype=np.float32)
+                grad_tensor.set(data, place)
+
+
+def get_numeric_gradient(scope,
+                         op,
+                         inputs,
+                         input_to_check,
+                         output_names,
+                         delta=0.005,
+                         in_place=False):
+
+    set_input(scope, op, inputs, core.CPUPlace())
+    op.infer_shape(scope)
+
+    tensor_to_check = scope.find_var(input_to_check).get_tensor()
+
+    def product(dim):
+        return reduce(lambda a, b: a * b, dim, 1)
+
+    ctx = core.DeviceContext.create(core.CPUPlace())
+
+    def get_output():
+        sum = 0.0
+        for output_name in output_names:
+            op.run(scope, ctx)
+            sum += np.array(scope.find_var(output_name).get_tensor()).sum()
+        return sum
+
+    tensor_to_check = scope.find_var(input_to_check).get_tensor()
+    tensor_size = product(tensor_to_check.get_dims())
+    gradient_flat = np.zeros(shape=(tensor_size, ), dtype='float32')
+    # we only compute gradient of one element each time.
+    # we use a for loop to compute the gradient of every element.
+    for i in xrange(tensor_size):
+        if in_place:
+            set_input(scope, op, inputs, core.CPUPlace())
+
+        # get one input element throw it's index i.
+        origin = tensor_to_check.get_float_element(i)
+        # add delta to it, run op and then get the sum of the result tensor.
+        x_pos = origin + delta
+        tensor_to_check.set_float_element(i, x_pos)
+        y_pos = get_output()
+
+        if in_place:
+            set_input(scope, op, inputs, core.CPUPlace())
+
+        x_neg = origin - delta
+        tensor_to_check.set_float_element(i, x_neg)
+        y_neg = get_output()
+
+        tensor_to_check.set_float_element(i, origin)
+        gradient_flat[i] = (y_pos - y_neg) / delta / 2
+
+    return gradient_flat.reshape(tensor_to_check.get_dims())
+
+
+def get_backward_op(scope, op, no_grad_set):
+    backward_op = core.Operator.backward(op, no_grad_set)
+    for input in backward_op.input_vars():
+        var = scope.new_var(input)
+        var.get_tensor()
+    for output in backward_op.output_vars():
+        var = scope.new_var(output)
+        var.get_tensor()
+    return backward_op
+
+
+def get_gradient(scope, op, inputs, outputs, grad_name, place,
+                 no_grad_set=None):
+    ctx = core.DeviceContext.create(place)
+
+    set_input(scope, op, inputs, place)
+
+    op.infer_shape(scope)
+    op.run(scope, ctx)
+
+    if no_grad_set is None:
+        no_grad_set = set()
+
+    backward_op = get_backward_op(scope, op, no_grad_set)
+    set_output_grad(scope, op, outputs, place)
+
+    backward_op.infer_shape(scope)
+    backward_op.run(scope, ctx)
+
+    out = np.array(scope.find_var(grad_name).get_tensor())
+    return out
+
+
+class OpTest(unittest.TestCase):
+    def check_output_with_place(self, place):
+        self.scope = core.Scope()
+        op_inputs = self.inputs if hasattr(self, "inputs") else dict()
+        op_outputs = self.outputs if hasattr(self, "outputs") else dict()
+        op_attrs = self.attrs if hasattr(self, "attrs") else dict()
+        self.op = create_op(self.scope, self.op_type, op_inputs, op_outputs,
+                            op_attrs)
+        if isinstance(place, core.GPUPlace) and not self.op.support_gpu():
+            return
+        set_input(self.scope, self.op, self.inputs, place)
+        self.op.infer_shape(self.scope)
+        ctx = core.DeviceContext.create(place)
+        self.op.run(self.scope, ctx)
+
+        for out_name, out_dup in Operator.get_op_outputs(self.op.type()):
+            if out_name not in self.outputs:
+                continue
+
+            if out_dup:
+                sub_out = self.outputs[out_name]
+                if not isinstance(sub_out, list):
+                    raise AssertionError("sub_out type %s is not list",
+                                         type(sub_out))
+
+                for sub_out_name, expect in sub_out:
+                    actual = np.array(
+                        self.scope.find_var(sub_out_name).get_tensor())
+                    self.assertTrue(
+                        np.allclose(
+                            actual, expect, atol=1e-05),
+                        "output name: " + out_name + " has diff")
+            else:
+                actual = np.array(self.scope.find_var(out_name).get_tensor())
+                expect = self.outputs[out_name]
+                self.assertTrue(
+                    np.allclose(
+                        actual, expect, atol=1e-05),
+                    "output name: " + out_name + " has diff")
+
+    def check_output(self):
+        places = [core.CPUPlace()]
+        if core.is_compile_gpu():
+            places.append(core.GPUPlace(0))
+        for place in places:
+            self.check_output_with_place(place)
+
+    def __assert_is_close(self, numeric_grads, analytic_grads, names,
+                          max_relative_error, msg_prefix):
+
+        for a, b, name in itertools.izip(numeric_grads, analytic_grads, names):
+            abs_a = np.abs(a)
+            abs_a[abs_a < 1e-3] = 1
+
+            diff_mat = np.abs(a - b) / abs_a
+            max_diff = np.max(diff_mat)
+
+            def err_msg():
+                offset = np.argmax(diff_mat > max_relative_error)
+                return "%s Variable %s max gradient diff %f over limit %f, the first " \
+                  "error element is %d" % (
+                   msg_prefix, name, max_diff, max_relative_error, offset)
+
+            self.assertLessEqual(max_diff, max_relative_error, err_msg())
+
+    def check_grad(self,
+                   inputs_to_check,
+                   output_names,
+                   no_grad_set=None,
+                   in_place=False,
+                   max_relative_error=0.005):
+        self.scope = core.Scope()
+        op_inputs = self.inputs if hasattr(self, "inputs") else dict()
+        op_outputs = self.outputs if hasattr(self, "outputs") else dict()
+        op_attrs = self.attrs if hasattr(self, "attrs") else dict()
+        self.op = create_op(self.scope, self.op_type, op_inputs, op_outputs,
+                            op_attrs)
+        if no_grad_set is None:
+            no_grad_set = set()
+
+        if not type(output_names) is list:
+            output_names = [output_names]
+
+        numeric_grads = [
+            get_numeric_gradient(
+                self.scope,
+                self.op,
+                self.inputs,
+                input_to_check,
+                output_names,
+                in_place=in_place) for input_to_check in inputs_to_check
+        ]
+        grad_names = [
+            grad_var_name(input_to_check) for input_to_check in inputs_to_check
+        ]
+
+        cpu_place = core.CPUPlace()
+        cpu_analytic_grads = [
+            get_gradient(self.scope, self.op, self.inputs, self.outputs,
+                         grad_name, cpu_place, no_grad_set)
+            for grad_name in grad_names
+        ]
+
+        self.__assert_is_close(numeric_grads, cpu_analytic_grads, grad_names,
+                               max_relative_error,
+                               "Gradient Check On %s" % str(cpu_place))
+
+        if core.is_compile_gpu() and self.op.support_gpu():
+            gpu_place = core.GPUPlace(0)
+            gpu_analytic_grads = [
+                get_gradient(self.scope, self.op, self.inputs, self.outputs,
+                             grad_name, gpu_place, no_grad_set)
+                for grad_name in grad_names
+            ]
+
+            self.__assert_is_close(numeric_grads, gpu_analytic_grads,
+                                   grad_names, max_relative_error,
+                                   "Gradient Check On %s" % str(gpu_place))
+
+            for c_grad, g_grad, name in itertools.izip(
+                    cpu_analytic_grads, gpu_analytic_grads, grad_names):
+                self.assertTrue(
+                    np.allclose(
+                        c_grad, g_grad, atol=1e-4),
+                    "output name: " + name + " has diff")
diff --git a/python/paddle/v2/framework/tests/op_test_util.py b/python/paddle/v2/framework/tests/op_test_util.py
deleted file mode 100644
index 370f27eaf6..0000000000
--- a/python/paddle/v2/framework/tests/op_test_util.py
+++ /dev/null
@@ -1,72 +0,0 @@
-import numpy
-import paddle.v2.framework.core as core
-from paddle.v2.framework.op import Operator
-
-
-class OpTestMeta(type):
-    """
-    Operator Test ClassMeta.
-
-    It injects `test_all` method into user's OperatorTest class, to make Python
-    unittest module run that method.
-
-    The `test_all` read what value is stored in `self`. It use self's values to
-    create and run a operator, and check whether that op is OK or not.
-
-    See `test_add_two_op` for example usage.
-    """
-
-    def __new__(cls, name, bases, attrs):
-        obj = super(OpTestMeta, cls).__new__(cls, name, bases, attrs)
-
-        def test_all(self):
-            scope = core.Scope()
-            kwargs = dict()
-            places = [core.CPUPlace()]
-            if core.is_compile_gpu():
-                places.append(core.GPUPlace(0))
-
-            for place in places:
-                for in_name in Operator.get_op_input_names(self.type):
-                    if hasattr(self, "inputs") and in_name in self.inputs:
-                        kwargs[in_name] = in_name
-                        var = scope.new_var(in_name).get_tensor()
-                        arr = self.inputs[in_name]
-                        var.set_dims(arr.shape)
-                        var.set(arr, place)
-                    else:
-                        kwargs[in_name] = "@EMPTY@"
-
-                for out_name in Operator.get_op_output_names(self.type):
-                    if not hasattr(self, "outputs"):
-                        raise ValueError(
-                            "The test op must set self.outputs dict.")
-                    if out_name not in self.outputs:
-                        raise ValueError("The %s is not in self.outputs dict." %
-                                         (out_name))
-                    kwargs[out_name] = out_name
-                    scope.new_var(out_name).get_tensor()
-
-                for attr_name in Operator.get_op_attr_names(self.type):
-                    if hasattr(self, "attrs") and attr_name in self.attrs:
-                        kwargs[attr_name] = self.attrs[attr_name]
-
-                op = Operator(self.type, **kwargs)
-                if isinstance(place, core.GPUPlace) and not op.support_gpu():
-                    return
-
-                op.infer_shape(scope)
-
-                ctx = core.DeviceContext.create(place)
-                op.run(scope, ctx)
-
-                for out_name in Operator.get_op_output_names(self.type):
-                    actual = numpy.array(scope.find_var(out_name).get_tensor())
-                    expect = self.outputs[out_name]
-                    self.assertTrue(
-                        numpy.allclose(
-                            actual, expect, atol=1e-05),
-                        "output name: " + out_name + " has diff")
-
-        obj.test_all = test_all
-        return obj
diff --git a/python/paddle/v2/framework/tests/test_accuracy_op.py b/python/paddle/v2/framework/tests/test_accuracy_op.py
new file mode 100644
index 0000000000..b6f3a35d6f
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_accuracy_op.py
@@ -0,0 +1,26 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestAccuracyOp(OpTest):
+    def setUp(self):
+        self.op_type = "accuracy"
+        n = 8192
+        infer = np.random.randint(0, 2, (n, 1)).astype("int")
+        label = np.random.randint(0, 2, (n, )).astype("int")
+        self.inputs = {'Inference': infer, "Label": label}
+        num_correct = 0
+        for rowid in xrange(n):
+            for ele in infer[rowid]:
+                if ele == label[rowid]:
+                    num_correct += 1
+                    break
+        self.outputs = {'Accuracy': [num_correct / float(n)]}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_add_op.py b/python/paddle/v2/framework/tests/test_add_op.py
new file mode 100644
index 0000000000..3ca34d9b9f
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_add_op.py
@@ -0,0 +1,20 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestAddOp(OpTest):
+    def setUp(self):
+        self.op_type = "add"
+        self.inputs = {
+            'X': np.random.random((102, 105)).astype("float32"),
+            'Y': np.random.random((102, 105)).astype("float32")
+        }
+        self.outputs = {'Out': self.inputs['X'] + self.inputs['Y']}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_add_two_op.py b/python/paddle/v2/framework/tests/test_add_two_op.py
deleted file mode 100644
index a578e74eca..0000000000
--- a/python/paddle/v2/framework/tests/test_add_two_op.py
+++ /dev/null
@@ -1,23 +0,0 @@
-import unittest
-
-import numpy
-import paddle.v2.framework.core as core
-from paddle.v2.framework.op import Operator
-
-from op_test_util import OpTestMeta
-
-
-class TestAddOp(unittest.TestCase):
-    __metaclass__ = OpTestMeta
-
-    def setUp(self):
-        self.type = "add"
-        self.inputs = {
-            'X': numpy.random.random((102, 105)).astype("float32"),
-            'Y': numpy.random.random((102, 105)).astype("float32")
-        }
-        self.outputs = {'Out': self.inputs['X'] + self.inputs['Y']}
-
-
-if __name__ == '__main__':
-    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_concat_op.py b/python/paddle/v2/framework/tests/test_concat_op.py
new file mode 100644
index 0000000000..656563f96e
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_concat_op.py
@@ -0,0 +1,22 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestConcatOp(OpTest):
+    def setUp(self):
+        self.op_type = "concat"
+        x0 = np.random.random((2, 3, 2, 5)).astype('float32')
+        x1 = np.random.random((2, 3, 3, 5)).astype('float32')
+        x2 = np.random.random((2, 3, 4, 5)).astype('float32')
+        axis = 2
+        self.inputs = {'X': [('x0', x0), ('x1', x1), ('x2', x2)]}
+        self.attrs = {'axis': axis}
+        self.outputs = {'Out': np.concatenate((x0, x1, x2), axis=axis)}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_cond_op.py b/python/paddle/v2/framework/tests/test_cond_op.py
new file mode 100644
index 0000000000..37177ae0b2
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_cond_op.py
@@ -0,0 +1,116 @@
+import logging
+import paddle.v2.framework.core as core
+import unittest
+import numpy as np
+from paddle.v2.framework.op import Operator, CondOp
+
+
+class PySimpleCond(object):
+    '''
+    A simple implementation of dynamic if-else based on numpy
+    '''
+
+    def __init__(self):
+        array = [1] * 10
+        for i in range(1, 10, 2):
+            array[i] = 0
+        self.cond = np.array(array)
+        self.x = np.ones(shape=(10, 1))
+
+    def forward(self):
+        self.index_t = np.where(self.cond == 1)
+        self.index_f = np.where(self.cond == 0)
+        y_t = self.x[self.index_t]
+        y_f = self.x[self.index_f]
+        y_t = y_t * 2.
+        y_f = y_f * (-2.)
+        output = np.zeros(shape=(10, 1))
+        output[self.index_t] = y_t
+        output[self.index_f] = y_f
+        return output
+
+
+class PySimpleCondTest(unittest.TestCase):
+    def setUp(self):
+        self.condnn = PySimpleCond()
+
+    def test_forward(self):
+        output = self.condnn.forward()
+
+
+def create_tensor(scope, name, shape, np_data):
+    tensor = scope.new_var(name).get_tensor()
+    tensor.set_dims(shape)
+    tensor.set(np_data, core.CPUPlace())
+    return tensor
+
+
+class TestCondOp(unittest.TestCase):
+    '''
+    Test CondOp
+
+    equation:
+        cond = [True, False, True, False, ...]
+        y[index_t] = x[index_t] * 2.
+        y[index_f] = x[index_f] * -2.
+    outputs:
+        y
+    '''
+
+    def setUp(self):
+        self.py_cond = PySimpleCond()
+
+    def forward(self):
+        self.scope = core.Scope()
+        self.create_global_variables()
+        self.create_cond_op()
+        self.create_sub_net()
+        ctx = core.DeviceContext.create(core.CPUPlace())
+        self.condop.infer_shape(self.scope)
+        self.condop.run(self.scope, ctx)
+        return np.array(self.scope.find_var("Out").get_tensor())
+
+    def create_global_variables(self):
+        x_np_data = self.py_cond.x
+        create_tensor(self.scope, "X", [10, 1], x_np_data)
+        cond_np_data = self.py_cond.cond.astype("int32")
+        create_tensor(self.scope, "cond", [10, 1], cond_np_data)
+        self.scope.new_var("SubScopes")
+        self.scope.new_var("IndexTensors")
+        self.scope.new_var("Out")
+
+    def create_cond_op(self):
+        self.condop = CondOp(
+            Cond="cond",
+            Xs=["X"],
+            Outs=["Out"],
+            SubScopes="SubScopes",
+            IndexTensors="IndexTensors")
+
+    def create_sub_net(self):
+        truenet = core.Net.create()
+        scale_op_t = Operator("scale", X='X', Out='Out', scale=2.)
+        truenet.append_op(scale_op_t)
+        truenet.complete_add_op(True)
+        self.condop.set_truenet(truenet)
+
+        falsenet = core.Net.create()
+        scale_op_t = Operator("scale", X='X', Out='Out', scale=-2.)
+        falsenet.append_op(scale_op_t)
+        falsenet.complete_add_op(True)
+        self.condop.set_falsenet(falsenet)
+
+    def test_forward(self):
+        print 'test cond op forward'
+        pd_output = self.forward()
+        py_output = self.py_cond.forward()
+        print 'pd_output', pd_output
+        print
+        print 'py_output', py_output
+        self.assertEqual(pd_output.shape, py_output.shape)
+        print 'test passed'
+        return 0
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_cos_sim_op.py b/python/paddle/v2/framework/tests/test_cos_sim_op.py
index 32013a7999..d314ce391e 100644
--- a/python/paddle/v2/framework/tests/test_cos_sim_op.py
+++ b/python/paddle/v2/framework/tests/test_cos_sim_op.py
@@ -1,17 +1,14 @@
 import unittest
 import numpy as np
-from gradient_checker import GradientChecker, create_op
-from op_test_util import OpTestMeta
+from op_test import OpTest
 
 
-class TestCosSimOp(unittest.TestCase):
-    __metaclass__ = OpTestMeta
-
+class TestCosSimOp(OpTest):
     def setUp(self):
-        self.type = "cos_sim"
+        self.op_type = "cos_sim"
         self.inputs = {
-            'X': np.random.random((32, 64)).astype("float32"),
-            'Y': np.random.random((32, 64)).astype("float32")
+            'X': np.random.random((6, 5)).astype("float32"),
+            'Y': np.random.random((6, 5)).astype("float32")
         }
         expect_x_norm = np.linalg.norm(self.inputs['X'], axis=1)
         expect_y_norm = np.linalg.norm(self.inputs['Y'], axis=1)
@@ -23,37 +20,73 @@ class TestCosSimOp(unittest.TestCase):
             'Out': np.expand_dims(expect_out, 1)
         }
 
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.05)
+
+    def test_check_grad_ingore_x(self):
+        self.check_grad(
+            ['Y'], 'Out', max_relative_error=0.05, no_grad_set=set("X"))
+
+    def test_check_grad_ingore_y(self):
+        self.check_grad(
+            ['X'], 'Out', max_relative_error=0.05, no_grad_set=set('Y'))
+
 
-class TestCosSimGradOp(GradientChecker):
+class TestCosSimOp2(TestCosSimOp):
     def setUp(self):
-        self.op = create_op("cos_sim")
+        self.op_type = "cos_sim"
         self.inputs = {
-            'X': np.random.random((10, 5)).astype("float32"),
-            'Y': np.random.random((10, 5)).astype("float32")
+            'X': np.random.random((6, 5)).astype("float32"),
+            'Y': np.random.random((1, 5)).astype("float32")
+        }
+        expect_x_norm = np.linalg.norm(self.inputs['X'], axis=1)
+        expect_y_norm = np.linalg.norm(self.inputs['Y'], axis=1)
+        expect_out = (self.inputs['X'] * self.inputs['Y']).sum(axis=1) / \
+            expect_x_norm / expect_y_norm
+        self.outputs = {
+            'XNorm': np.expand_dims(expect_x_norm, 1),
+            'YNorm': np.expand_dims(expect_y_norm, 1),
+            'Out': np.expand_dims(expect_out, 1)
         }
 
-    def test_cpu_gpu_compare(self):
-        self.compare_grad(self.op, self.inputs)
 
-    def test_normal(self):
-        self.check_grad(
-            self.op, self.inputs, ["X", "Y"], "Out", max_relative_error=0.05)
+class TestCosSimOp3(TestCosSimOp):
+    def setUp(self):
+        self.op_type = "cos_sim"
+        self.inputs = {
+            'X': np.random.random((6, 5, 2)).astype("float32"),
+            'Y': np.random.random((6, 5, 2)).astype("float32")
+        }
+        expect_x_norm = np.linalg.norm(self.inputs['X'], axis=(1, 2))
+        expect_y_norm = np.linalg.norm(self.inputs['Y'], axis=(1, 2))
+        expect_out = (self.inputs['X'] * self.inputs['Y']).sum(axis=(1, 2)) / \
+            expect_x_norm / expect_y_norm
+        self.outputs = {
+            'XNorm': np.expand_dims(expect_x_norm, 1),
+            'YNorm': np.expand_dims(expect_y_norm, 1),
+            'Out': np.expand_dims(expect_out, 1)
+        }
 
-    def test_ignore_x(self):
-        self.check_grad(
-            self.op,
-            self.inputs, ["Y"],
-            "Out",
-            max_relative_error=0.05,
-            no_grad_set={"X"})
 
-    def test_ignore_y(self):
-        self.check_grad(
-            self.op,
-            self.inputs, ["X"],
-            "Out",
-            max_relative_error=0.05,
-            no_grad_set={"Y"})
+class TestCosSimOp4(TestCosSimOp):
+    def setUp(self):
+        self.op_type = "cos_sim"
+        self.inputs = {
+            'X': np.random.random((6, 5, 2)).astype("float32"),
+            'Y': np.random.random((1, 5, 2)).astype("float32")
+        }
+        expect_x_norm = np.linalg.norm(self.inputs['X'], axis=(1, 2))
+        expect_y_norm = np.linalg.norm(self.inputs['Y'], axis=(1, 2))
+        expect_out = (self.inputs['X'] * self.inputs['Y']).sum(axis=(1, 2)) / \
+            expect_x_norm / expect_y_norm
+        self.outputs = {
+            'XNorm': np.expand_dims(expect_x_norm, 1),
+            'YNorm': np.expand_dims(expect_y_norm, 1),
+            'Out': np.expand_dims(expect_out, 1)
+        }
 
 
 if __name__ == '__main__':
diff --git a/python/paddle/v2/framework/tests/test_cross_entropy_op.py b/python/paddle/v2/framework/tests/test_cross_entropy_op.py
deleted file mode 100644
index d4277f2a42..0000000000
--- a/python/paddle/v2/framework/tests/test_cross_entropy_op.py
+++ /dev/null
@@ -1,37 +0,0 @@
-import unittest
-import numpy
-from op_test_util import OpTestMeta
-from gradient_checker import GradientChecker, create_op
-
-
-class TestCrossEntropy(unittest.TestCase):
-    __metaclass__ = OpTestMeta
-
-    def setUp(self):
-        self.type = "onehot_cross_entropy"
-        batch_size = 30
-        class_num = 10
-        X = numpy.random.random((batch_size, class_num)).astype("float32")
-        label = 5 * numpy.ones(batch_size).astype("int32")
-        self.inputs = {'X': X, 'label': label}
-        Y = []
-        for i in range(0, batch_size):
-            Y.append(-numpy.log(X[i][label[i]]))
-        self.outputs = {'Y': numpy.array(Y).astype("float32")}
-
-
-class CrossEntropyGradOpTest(GradientChecker):
-    def test_check_grad(self):
-        op = create_op("onehot_cross_entropy")
-        batch_size = 30
-        class_num = 10
-        inputs = {
-            "X": numpy.random.uniform(
-                0.1, 1.0, [batch_size, class_num]).astype("float32"),
-            "label": (class_num / 2) * numpy.ones(batch_size).astype("int32")
-        }
-        self.check_grad(op, inputs, set("X"), "Y")
-
-
-if __name__ == "__main__":
-    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_elementwise_mul_op.py b/python/paddle/v2/framework/tests/test_elementwise_mul_op.py
new file mode 100644
index 0000000000..e268cfddb2
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_elementwise_mul_op.py
@@ -0,0 +1,157 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestElementwiseMulOp_Matrix(OpTest):
+    def setUp(self):
+        self.op_type = "elementwise_mul"
+        """ Warning
+        CPU gradient check error!
+        'X': np.random.random((32,84)).astype("float32"),
+        'Y': np.random.random((32,84)).astype("float32")
+        """
+        self.inputs = {
+            'X': np.random.uniform(0.1, 1, [13, 17]).astype("float32"),
+            'Y': np.random.uniform(0.1, 1, [13, 17]).astype("float32")
+        }
+        self.outputs = {'Out': np.multiply(self.inputs['X'], self.inputs['Y'])}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.1)
+
+    def test_check_grad_ingore_x(self):
+        self.check_grad(
+            ['Y'], 'Out', max_relative_error=0.1, no_grad_set=set("X"))
+
+    def test_check_grad_ingore_y(self):
+        self.check_grad(
+            ['X'], 'Out', max_relative_error=0.1, no_grad_set=set('Y'))
+
+
+class TestElementwiseMulOp_Vector(OpTest):
+    def setUp(self):
+        self.op_type = "elementwise_mul"
+        self.inputs = {
+            'X': np.random.random((32, )).astype("float32"),
+            'Y': np.random.random((32, )).astype("float32")
+        }
+        self.outputs = {'Out': np.multiply(self.inputs['X'], self.inputs['Y'])}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.1)
+
+    def test_check_grad_ingore_x(self):
+        self.check_grad(
+            ['Y'], 'Out', max_relative_error=0.1, no_grad_set=set("X"))
+
+    def test_check_grad_ingore_y(self):
+        self.check_grad(
+            ['X'], 'Out', max_relative_error=0.1, no_grad_set=set('Y'))
+
+
+class TestElementwiseMulOp_broadcast_0(OpTest):
+    def setUp(self):
+        self.op_type = "elementwise_mul"
+        self.inputs = {
+            'X': np.random.rand(2, 3, 4).astype(np.float32),
+            'Y': np.random.rand(2).astype(np.float32)
+        }
+
+        self.attrs = {'axis': 0}
+        self.outputs = {
+            'Out': self.inputs['X'] * self.inputs['Y'].reshape(2, 1, 1)
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.1)
+
+    def test_check_grad_ingore_x(self):
+        self.check_grad(
+            ['Y'], 'Out', max_relative_error=0.1, no_grad_set=set("X"))
+
+    def test_check_grad_ingore_y(self):
+        self.check_grad(
+            ['X'], 'Out', max_relative_error=0.1, no_grad_set=set('Y'))
+
+
+class TestElementwiseMulOp_broadcast_1(OpTest):
+    def setUp(self):
+        self.op_type = "elementwise_mul"
+        self.inputs = {
+            'X': np.random.rand(2, 3, 4).astype(np.float32),
+            'Y': np.random.rand(3).astype(np.float32)
+        }
+
+        self.attrs = {'axis': 1}
+        self.outputs = {
+            'Out': self.inputs['X'] * self.inputs['Y'].reshape(1, 3, 1)
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.1)
+
+    def test_check_grad_ingore_x(self):
+        self.check_grad(
+            ['Y'], 'Out', max_relative_error=0.1, no_grad_set=set("X"))
+
+    def test_check_grad_ingore_y(self):
+        self.check_grad(
+            ['X'], 'Out', max_relative_error=0.1, no_grad_set=set('Y'))
+
+
+class TestElementwiseMulOp_broadcast_2(OpTest):
+    def setUp(self):
+        self.op_type = "elementwise_mul"
+        self.inputs = {
+            'X': np.random.rand(2, 3, 4).astype(np.float32),
+            'Y': np.random.rand(4).astype(np.float32)
+        }
+
+        self.outputs = {
+            'Out': self.inputs['X'] * self.inputs['Y'].reshape(1, 1, 4)
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.1)
+
+    def test_check_grad_ingore_x(self):
+        self.check_grad(
+            ['Y'], 'Out', max_relative_error=0.1, no_grad_set=set("X"))
+
+    def test_check_grad_ingore_y(self):
+        self.check_grad(
+            ['X'], 'Out', max_relative_error=0.1, no_grad_set=set('Y'))
+
+
+class TestElementwiseMulOp_broadcast_3(OpTest):
+    def setUp(self):
+        self.op_type = "elementwise_mul"
+        self.inputs = {
+            'X': np.random.rand(2, 3, 4, 5).astype(np.float32),
+            'Y': np.random.rand(3, 4).astype(np.float32)
+        }
+
+        self.attrs = {'axis': 1}
+        self.outputs = {
+            'Out': self.inputs['X'] * self.inputs['Y'].reshape(1, 3, 4, 1)
+        }
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_fc_op.py b/python/paddle/v2/framework/tests/test_fc_op.py
new file mode 100644
index 0000000000..9f56fe5049
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_fc_op.py
@@ -0,0 +1,62 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestFCOp1(OpTest):
+    def setUp(self):
+        x0 = np.random.random((16, 32)).astype("float32")
+        w0 = np.random.random((32, 10)).astype("float32")
+
+        mul_out0 = np.dot(x0, w0)
+        identity_out = mul_out0
+
+        self.op_type = "fc"
+        self.inputs = {"X": [("X0", x0)], "W": [("W0", w0)]}
+        self.outputs = {"MulOut": [("MulOut0", mul_out0)], "Out": identity_out}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(["X0", "W0"], "Out", max_relative_error=0.01)
+
+
+class TestFCOp2(OpTest):
+    def setUp(self):
+        x0 = np.random.random((16, 4, 8)).astype("float32")
+        x1 = np.random.random((4, 4, 32)).astype("float32")
+        w0 = np.random.random((32, 10)).astype("float32")
+        w1 = np.random.random((32, 10)).astype("float32")
+        b = np.random.random(10).astype("float32")
+
+        mul_out0 = np.dot(x0.reshape(16, 4 * 8), w0)
+        mul_out1 = np.dot(x1.reshape(4 * 4, 32), w1)
+        sum_out = mul_out0 + mul_out1
+        add_out = np.add(sum_out, b)
+        sigmoid_out = 1 / (1 + np.exp(-add_out))
+
+        self.op_type = "fc"
+        self.inputs = {
+            "X": [("X0", x0), ("X1", x1)],
+            "W": [("W0", w0), ("W1", w1)],
+            "B": b
+        }
+        self.attrs = {"xNumColDims": [1, 2], "activation": "sigmoid"}
+        self.outputs = {
+            "MulOut": [("MulOut0", mul_out0), ("MulOut1", mul_out1)],
+            "SumOut": sum_out,
+            "AddOut": add_out,
+            "Out": sigmoid_out
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(
+            ["X0", "X1", "W0", "W1", "B"], "Out", max_relative_error=0.01)
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_fill_zeros_like_op.py b/python/paddle/v2/framework/tests/test_fill_zeros_like_op.py
index e5c862605f..2473daaba2 100644
--- a/python/paddle/v2/framework/tests/test_fill_zeros_like_op.py
+++ b/python/paddle/v2/framework/tests/test_fill_zeros_like_op.py
@@ -1,16 +1,17 @@
 import unittest
-from op_test_util import OpTestMeta
-import numpy
+import numpy as np
+from op_test import OpTest
 
 
-class TestFillZerosLikeOp(unittest.TestCase):
-    __metaclass__ = OpTestMeta
-
+class TestFillZerosLikeOp(OpTest):
     def setUp(self):
-        self.type = "fill_zeros_like"
-        self.inputs = {'Src': numpy.random.random((219, 232)).astype("float32")}
-        self.outputs = {'Dst': numpy.zeros_like(self.inputs['Src'])}
+        self.op_type = "fill_zeros_like"
+        self.inputs = {'Src': np.random.random((219, 232)).astype("float32")}
+        self.outputs = {'Dst': np.zeros_like(self.inputs["Src"])}
+
+    def test_check_output(self):
+        self.check_output()
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_gather_op.py b/python/paddle/v2/framework/tests/test_gather_op.py
index e3de3fd0a1..b0ab429ef1 100644
--- a/python/paddle/v2/framework/tests/test_gather_op.py
+++ b/python/paddle/v2/framework/tests/test_gather_op.py
@@ -1,30 +1,20 @@
 import unittest
-from op_test_util import OpTestMeta
-from gradient_checker import GradientChecker, create_op
-import numpy
-import paddle.v2.framework.core as core
-from paddle.v2.framework.op import Operator
+import numpy as np
+from op_test import OpTest
 
 
-class TestGatherOp(unittest.TestCase):
-    __metaclass__ = OpTestMeta
-
+class TestGatherOp(OpTest):
     def setUp(self):
-        self.type = "gather"
-        xnp = numpy.random.random((10, 20)).astype("float32")
-        self.inputs = {
-            'X': xnp,
-            'Index': numpy.array([1, 3, 5]).astype("int32")
-        }
-        self.outputs = {'Out': self.inputs['X'][self.inputs['Index']]}
+        self.op_type = "gather"
+        xnp = np.random.random((10, 20)).astype("float32")
+        self.inputs = {'X': xnp, 'Index': np.array([1, 3, 5]).astype("int32")}
+        self.outputs = {'Out': self.inputs["X"][self.inputs["Index"]]}
 
+    def test_check_output(self):
+        self.check_output()
 
-class TestGatherGradOp(GradientChecker):
-    def test_gather_grad(self):
-        op = create_op("gather")
-        xnp = numpy.random.random((10, 20)).astype("float32")
-        inputs = {'X': xnp, 'Index': numpy.array([1, 3, 5]).astype("int32")}
-        self.check_grad(op, inputs, set("X"), "Out")
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Out')
 
 
 if __name__ == "__main__":
diff --git a/python/paddle/v2/framework/tests/test_gaussian_random_op.py b/python/paddle/v2/framework/tests/test_gaussian_random_op.py
index f95ed70b58..1888ee28f9 100644
--- a/python/paddle/v2/framework/tests/test_gaussian_random_op.py
+++ b/python/paddle/v2/framework/tests/test_gaussian_random_op.py
@@ -4,7 +4,7 @@ from paddle.v2.framework.op import Operator
 import numpy
 
 
-class GaussianRandomTest(unittest.TestCase):
+class TestGaussianRandomOp(unittest.TestCase):
     def test_cpu(self):
         self.gaussian_random_test(place=core.CPUPlace())
 
@@ -14,11 +14,11 @@ class GaussianRandomTest(unittest.TestCase):
 
     def gaussian_random_test(self, place):
         scope = core.Scope()
-        scope.new_var("Out").get_tensor()
+        scope.new_var('Out').get_tensor()
 
         op = Operator(
             "gaussian_random",
-            Out="Out",
+            Out='Out',
             dims=[1000, 784],
             mean=.0,
             std=1.,
@@ -27,10 +27,10 @@ class GaussianRandomTest(unittest.TestCase):
         op.infer_shape(scope)
         context = core.DeviceContext.create(place)
         op.run(scope, context)
-        tensor = numpy.array(scope.find_var("Out").get_tensor())
+        tensor = numpy.array(scope.find_var('Out').get_tensor())
         self.assertAlmostEqual(numpy.mean(tensor), .0, delta=0.1)
         self.assertAlmostEqual(numpy.std(tensor), 1., delta=0.1)
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_gradient_checker.py b/python/paddle/v2/framework/tests/test_gradient_checker.py
index e8a7f848df..85117bf960 100644
--- a/python/paddle/v2/framework/tests/test_gradient_checker.py
+++ b/python/paddle/v2/framework/tests/test_gradient_checker.py
@@ -1,42 +1,45 @@
 import unittest
-import numpy
-from paddle.v2.framework.op import Operator
-from gradient_checker import GradientChecker
-from gradient_checker import get_numeric_gradient
+import numpy as np
+import paddle.v2.framework.core as core
+from op_test import get_numeric_gradient
+from op_test import create_op
 
 
 class GetNumericGradientTest(unittest.TestCase):
     def test_add_op(self):
-        add_op = Operator("add", X="X", Y="Y", Out="Z")
-        x = numpy.random.random((10, 1)).astype("float32")
-        y = numpy.random.random((10, 1)).astype("float32")
-
-        arr = get_numeric_gradient(add_op, {"X": x, "Y": y}, "Z", "X")
+        x = np.random.random((10, 1)).astype("float32")
+        y = np.random.random((10, 1)).astype("float32")
+        z = x + y
+        scope = core.Scope()
+        add_op = create_op(scope, "add", {'X': x, 'Y': y}, {'Out': z}, dict())
+        arr = get_numeric_gradient(scope, add_op, {'X': x,
+                                                   'Y': y}, 'X', ['Out'])
         self.assertAlmostEqual(arr.mean(), 1.0, delta=1e-4)
 
     def test_softmax_op(self):
         def stable_softmax(x):
             """Compute the softmax of vector x in a numerically stable way."""
-            shiftx = x - numpy.max(x)
-            exps = numpy.exp(shiftx)
-            return exps / numpy.sum(exps)
+            shiftx = x - np.max(x)
+            exps = np.exp(shiftx)
+            return exps / np.sum(exps)
 
         def label_softmax_grad(Y, dY):
             dX = Y * 0.0
             for i in range(Y.shape[0]):
-                d = numpy.dot(Y[i, :], dY[i, :])
+                d = np.dot(Y[i, :], dY[i, :])
                 dX[i, :] = Y[i, :] * (dY[i, :] - d)
             return dX
 
-        softmax_op = Operator("softmax", X="X", Y="Y")
-
-        X = numpy.random.random((2, 2)).astype("float32")
-        Y = numpy.apply_along_axis(stable_softmax, 1, X)
-        dY = numpy.ones(Y.shape)
+        X = np.random.random((2, 2)).astype("float32")
+        Y = np.apply_along_axis(stable_softmax, 1, X)
+        dY = np.ones(Y.shape)
         dX = label_softmax_grad(Y, dY)
 
-        arr = get_numeric_gradient(softmax_op, {"X": X}, "Y", "X")
-        numpy.testing.assert_almost_equal(arr, dX, decimal=1e-2)
+        scope = core.Scope()
+        softmax_op = create_op(scope, "softmax", {"X": X}, {"Y": Y}, dict())
+
+        arr = get_numeric_gradient(scope, softmax_op, {"X": X}, "X", "Y")
+        np.testing.assert_almost_equal(arr, dX, decimal=1e-2)
 
 
 if __name__ == "__main__":
diff --git a/python/paddle/v2/framework/tests/test_identity_op.py b/python/paddle/v2/framework/tests/test_identity_op.py
new file mode 100644
index 0000000000..26cec1fcc3
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_identity_op.py
@@ -0,0 +1,20 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestIdentityOp(OpTest):
+    def setUp(self):
+        self.op_type = "identity"
+        self.inputs = {'X': np.random.random((10, 10)).astype("float32")}
+        self.outputs = {'Y': self.inputs['X']}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Y')
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_lookup_table.py b/python/paddle/v2/framework/tests/test_lookup_table.py
deleted file mode 100644
index 4b7ce92c0f..0000000000
--- a/python/paddle/v2/framework/tests/test_lookup_table.py
+++ /dev/null
@@ -1,31 +0,0 @@
-import unittest
-import numpy as np
-from op_test_util import OpTestMeta
-from gradient_checker import GradientChecker, create_op
-
-
-class TestLookupTableOp(unittest.TestCase):
-    __metaclass__ = OpTestMeta
-
-    def setUp(self):
-        self.type = 'lookup_table'
-        table = np.random.random((17, 31)).astype('float32')
-        ids = np.random.randint(0, 17, 4).astype('int32')
-        self.inputs = {'W': table, 'Ids': ids}
-        self.outputs = {'Out': table[ids]}
-
-
-class TestLookupTableGradOp(GradientChecker):
-    def test_grad(self):
-        op = create_op('lookup_table')
-        table = np.random.random((17, 31)).astype('float32')
-        ids = np.random.randint(0, 17, 4).astype('int32')
-        inputs = {'W': table, 'Ids': ids}
-        # comapre gradients 
-        self.compare_grad(op, inputs, set(['Ids']))
-        # check gradients 
-        self.check_grad(op, inputs, set('W'), 'Out')
-
-
-if __name__ == '__main__':
-    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_lookup_table_op.py b/python/paddle/v2/framework/tests/test_lookup_table_op.py
new file mode 100644
index 0000000000..b259bb67e8
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_lookup_table_op.py
@@ -0,0 +1,22 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestLookupTableOp(OpTest):
+    def setUp(self):
+        self.op_type = "lookup_table"
+        table = np.random.random((17, 31)).astype("float32")
+        ids = np.random.randint(0, 17, 4).astype("int32")
+        self.inputs = {'W': table, 'Ids': ids}
+        self.outputs = {'Out': table[ids]}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['W'], 'Out', no_grad_set=set('Ids'))
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_mean_op.py b/python/paddle/v2/framework/tests/test_mean_op.py
index f32b3160d6..7823abd8f8 100644
--- a/python/paddle/v2/framework/tests/test_mean_op.py
+++ b/python/paddle/v2/framework/tests/test_mean_op.py
@@ -1,24 +1,20 @@
 import unittest
-from op_test_util import OpTestMeta
-from gradient_checker import GradientChecker, create_op
 import numpy as np
+from op_test import OpTest
 
 
-class TestMeanOp(unittest.TestCase):
-    __metaclass__ = OpTestMeta
-
+class TestMeanOp(OpTest):
     def setUp(self):
-        self.type = "mean"
-        self.inputs = {'X': np.random.random((32, 784)).astype("float32")}
-        self.outputs = {'Out': np.mean(self.inputs['X'])}
+        self.op_type = "mean"
+        self.inputs = {'X': np.random.random((10, 10)).astype("float32")}
+        self.outputs = {'Out': np.mean(self.inputs["X"])}
 
+    def test_check_output(self):
+        self.check_output()
 
-class MeanGradOpTest(GradientChecker):
-    def test_normal(self):
-        op = create_op("mean")
-        inputs = {"X": np.random.random((10, 10)).astype("float32")}
-        self.check_grad(op, inputs, set("X"), "Out")
+    def test_checkout_grad(self):
+        self.check_grad(['X'], 'Out')
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_minus_op.py b/python/paddle/v2/framework/tests/test_minus_op.py
index 5abdd4a69b..c56d7cb548 100644
--- a/python/paddle/v2/framework/tests/test_minus_op.py
+++ b/python/paddle/v2/framework/tests/test_minus_op.py
@@ -1,30 +1,23 @@
 import unittest
 import numpy as np
-from gradient_checker import GradientChecker, create_op
-from op_test_util import OpTestMeta
+from op_test import OpTest
 
 
-class MinusOpTest(unittest.TestCase):
-    __metaclass__ = OpTestMeta
-
+class TestMinusOp(OpTest):
     def setUp(self):
-        self.type = "minus"
+        self.op_type = "minus"
         self.inputs = {
             'X': np.random.random((32, 84)).astype("float32"),
             'Y': np.random.random((32, 84)).astype("float32")
         }
         self.outputs = {'Out': (self.inputs['X'] - self.inputs['Y'])}
 
+    def test_check_output(self):
+        self.check_output()
 
-class MinusGradTest(GradientChecker):
-    def test_left(self):
-        op = create_op("minus")
-        inputs = {
-            "X": np.random.random((10, 10)).astype("float32"),
-            "Y": np.random.random((10, 10)).astype("float32")
-        }
-        self.check_grad(op, inputs, ["X", 'Y'], "Out")
+    def test_check_grad(self):
+        self.check_grad(['X', 'Y'], 'Out')
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     unittest.main()
diff --git a/python/paddle/v2/framework/tests/mnist.py b/python/paddle/v2/framework/tests/test_mnist.py
similarity index 100%
rename from python/paddle/v2/framework/tests/mnist.py
rename to python/paddle/v2/framework/tests/test_mnist.py
diff --git a/python/paddle/v2/framework/tests/test_mul_op.py b/python/paddle/v2/framework/tests/test_mul_op.py
index 8c827e242e..b3d95a56b8 100644
--- a/python/paddle/v2/framework/tests/test_mul_op.py
+++ b/python/paddle/v2/framework/tests/test_mul_op.py
@@ -1,27 +1,35 @@
 import unittest
 import numpy as np
-from gradient_checker import GradientChecker, create_op
-from op_test_util import OpTestMeta
-from paddle.v2.framework.op import Operator
+from op_test import OpTest
 
 
-class TestMulOp(unittest.TestCase):
-    __metaclass__ = OpTestMeta
-
+class TestMulOp(OpTest):
     def setUp(self):
-        self.type = "mul"
+        self.op_type = "mul"
         self.inputs = {
             'X': np.random.random((32, 84)).astype("float32"),
             'Y': np.random.random((84, 100)).astype("float32")
         }
         self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])}
 
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5)
 
-class TestMulOp2(unittest.TestCase):
-    __metaclass__ = OpTestMeta
+    def test_check_grad_ingore_x(self):
+        self.check_grad(
+            ['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X"))
 
+    def test_check_grad_ingore_y(self):
+        self.check_grad(
+            ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y'))
+
+
+class TestMulOp2(OpTest):
     def setUp(self):
-        self.type = "mul"
+        self.op_type = "mul"
         self.inputs = {
             'X': np.random.random((15, 4, 12, 10)).astype("float32"),
             'Y': np.random.random((4, 30, 8, 2, 9)).astype("float32")
@@ -32,72 +40,20 @@ class TestMulOp2(unittest.TestCase):
                           self.inputs['Y'].reshape(4 * 30, 8 * 2 * 9))
         }
 
+    def test_check_output(self):
+        self.check_output()
 
-class TestMulGradOp(GradientChecker):
-    def setUp(self):
-        self.op = create_op("mul")
-        self.inputs = {
-            'X': np.random.random((32, 84)).astype("float32"),
-            'Y': np.random.random((84, 100)).astype("float32")
-        }
-
-    def test_cpu_gpu_compare(self):
-        self.compare_grad(self.op, self.inputs)
-
-    def test_normal(self):
-        # mul op will enlarge the relative error
-        self.check_grad(
-            self.op, self.inputs, ["X", "Y"], "Out", max_relative_error=0.5)
-
-    def test_ignore_x(self):
-        self.check_grad(
-            self.op,
-            self.inputs, ["Y"],
-            "Out",
-            max_relative_error=0.5,
-            no_grad_set={"X"})
-
-    def test_ignore_y(self):
-        self.check_grad(
-            self.op,
-            self.inputs, ["X"],
-            "Out",
-            max_relative_error=0.5,
-            no_grad_set={"Y"})
-
-
-class TestMulGradTest2(GradientChecker):
-    def setUp(self):
-        self.op = Operator(
-            "mul", X="X", Y="Y", Out="Out", x_num_col_dims=2, y_num_col_dims=2)
-        self.inputs = {
-            "X": np.random.random((15, 4, 12, 10)).astype("float32"),
-            "Y": np.random.random((4, 30, 8, 2, 9)).astype("float32")
-        }
-
-    def test_cpu_gpu_compare(self):
-        self.compare_grad(self.op, self.inputs)
-
-    def test_normal(self):
-        self.check_grad(
-            self.op, self.inputs, ["X", "Y"], "Out", max_relative_error=0.5)
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5)
 
-    def test_ignore_x(self):
+    def test_check_grad_ingore_x(self):
         self.check_grad(
-            self.op,
-            self.inputs, ["Y"],
-            "Out",
-            max_relative_error=0.5,
-            no_grad_set={"X"})
+            ['Y'], 'Out', max_relative_error=0.5, no_grad_set=set('X'))
 
-    def test_ignore_y(self):
+    def test_check_grad_ignore_y(self):
         self.check_grad(
-            self.op,
-            self.inputs, ["X"],
-            "Out",
-            max_relative_error=0.5,
-            no_grad_set={"Y"})
+            ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y'))
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_net.py b/python/paddle/v2/framework/tests/test_net.py
index e4b7cd480c..50cfb855f2 100644
--- a/python/paddle/v2/framework/tests/test_net.py
+++ b/python/paddle/v2/framework/tests/test_net.py
@@ -35,5 +35,5 @@ Op(plain_net), inputs:{all[W, X, Y]}, outputs:{all[Out, fc.out, pre_activation]}
         self.assertEqual(expected, "\n" + str(net))
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_onehot_cross_entropy_op.py b/python/paddle/v2/framework/tests/test_onehot_cross_entropy_op.py
new file mode 100644
index 0000000000..fd3cbdb803
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_onehot_cross_entropy_op.py
@@ -0,0 +1,30 @@
+import unittest
+import numpy
+from op_test import OpTest
+
+
+class TestOnehotCrossEntropyOp(OpTest):
+    def setUp(self):
+        self.op_type = "onehot_cross_entropy"
+        batch_size = 30
+        class_num = 10
+
+        X = numpy.random.uniform(0.1, 1.0,
+                                 [batch_size, class_num]).astype("float32")
+        labels = numpy.random.randint(0, class_num, batch_size, dtype="int32")
+
+        cross_entropy = numpy.asmatrix(
+            [[-numpy.log(X[i][labels[i]])] for i in range(X.shape[0])],
+            dtype="float32")
+        self.inputs = {"X": X, "label": labels}
+        self.outputs = {"Y": cross_entropy}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(["X"], "Y")
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_pad_op.py b/python/paddle/v2/framework/tests/test_pad_op.py
new file mode 100644
index 0000000000..9052e63b56
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_pad_op.py
@@ -0,0 +1,55 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestPadOp(OpTest):
+    def setUp(self):
+        self.initTestCase()
+        self.op_type = "pad"
+        self.inputs = {'X': np.random.random(self.shape).astype("float32"), }
+        self.attrs = {}
+        self.attrs['paddings'] = np.array(self.paddings).flatten()
+        self.attrs['pad_value'] = self.pad_value
+        self.outputs = {
+            'Out': np.pad(self.inputs['X'],
+                          self.paddings,
+                          mode='constant',
+                          constant_values=self.pad_value)
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X'], 'Out', max_relative_error=0.006)
+
+    def initTestCase(self):
+        self.shape = (16, 16)
+        self.paddings = [(0, 1), (2, 3)]
+        self.pad_value = 0
+
+
+class TestCase1(TestPadOp):
+    def initTestCase(self):
+        self.shape = (2, 3, 4, 4)
+        self.paddings = [(0, 1), (2, 3), (2, 1), (1, 1)]
+        self.pad_value = 0.5
+
+
+class TestCase2(TestPadOp):
+    def initTestCase(self):
+        self.shape = (2, 2, 2)
+        self.paddings = [(0, 0), (0, 0), (1, 2)]
+        self.pad_value = 1
+
+
+class TestCase3(TestPadOp):
+    def initTestCase(self):
+        self.shape = (8)
+        self.paddings = [(0, 1)]
+        self.pad_value = 0.9
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_reshape_op.py b/python/paddle/v2/framework/tests/test_reshape_op.py
new file mode 100644
index 0000000000..16bb6bb2af
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_reshape_op.py
@@ -0,0 +1,21 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestReshapeOp(OpTest):
+    def setUp(self):
+        self.op_type = "reshape"
+        self.inputs = {'X': np.random.random((10, 20)).astype("float32")}
+        self.attrs = {'shape': [10 * 20]}
+        self.outputs = {'Out': self.inputs['X'].reshape(self.attrs['shape'])}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(["X"], "Out")
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_rowwise_add_op.py b/python/paddle/v2/framework/tests/test_rowwise_add_op.py
index 8378c1cd21..336645bd99 100644
--- a/python/paddle/v2/framework/tests/test_rowwise_add_op.py
+++ b/python/paddle/v2/framework/tests/test_rowwise_add_op.py
@@ -1,68 +1,51 @@
 import unittest
 import numpy as np
-from op_test_util import OpTestMeta
-from gradient_checker import GradientChecker, create_op
+from op_test import OpTest
 
 
-class TestRowwiseAddOp(unittest.TestCase):
-    __metaclass__ = OpTestMeta
-
-    def setUp(self):
-        self.type = "rowwise_add"
-        self.inputs = {
-            'X': np.random.random((32, 84)).astype("float32"),
-            'b': np.random.random(84).astype("float32")
-        }
-        self.outputs = {'Out': np.add(self.inputs['X'], self.inputs['b'])}
-
-
-class TestRowwiseAddOp2(unittest.TestCase):
-    __metaclass__ = OpTestMeta
-
+class TestRowwiseAddOp(OpTest):
     def setUp(self):
-        self.type = "rowwise_add"
+        self.op_type = "rowwise_add"
         self.inputs = {
-            'X': np.random.random((13, 6, 7, 8)).astype("float32"),
-            'b': np.random.random((7, 8)).astype("float32")
+            'X': np.random.uniform(0.1, 1, [5, 10]).astype("float32"),
+            'b': np.random.uniform(0.1, 1, [10]).astype("float32")
         }
         self.outputs = {'Out': np.add(self.inputs['X'], self.inputs['b'])}
 
+    def test_check_output(self):
+        self.check_output()
 
-class TestRowwiseAddGradOp(GradientChecker):
-    def setUp(self):
-        self.op = create_op("rowwise_add")
-        self.inputs = {
-            "X": np.random.uniform(0.1, 1, [5, 10]).astype("float32"),
-            "b": np.random.uniform(0.1, 1, [10]).astype("float32")
-        }
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'b'], 'Out')
 
-    def test_normal(self):
-        self.check_grad(self.op, self.inputs, ["X", "b"], "Out")
+    def test_check_grad_ingore_b(self):
+        self.check_grad(['X'], 'Out', no_grad_set=set('b'))
 
-    def test_ignore_b(self):
-        self.check_grad(self.op, self.inputs, ["X"], "Out", no_grad_set={"b"})
+    def test_check_grad_ingore_x(self):
+        self.check_grad(['b'], 'Out', no_grad_set=set('X'))
 
-    def test_ignore_x(self):
-        self.check_grad(self.op, self.inputs, ["b"], "Out", no_grad_set={"X"})
 
-
-class TestRowwiseAddGradOp2(GradientChecker):
+class TestRowwiseAddOp2(OpTest):
     def setUp(self):
-        self.op = create_op("rowwise_add")
+        self.op_type = "rowwise_add"
         self.inputs = {
-            "X": np.random.uniform(0.1, 1, [2, 3, 2, 5]).astype("float32"),
-            "b": np.random.uniform(0.1, 1, [2, 5]).astype("float32")
+            'X': np.random.uniform(0.1, 1, [2, 3, 2, 5]).astype("float32"),
+            'b': np.random.uniform(0.1, 1, [2, 5]).astype("float32")
         }
+        self.outputs = {'Out': np.add(self.inputs['X'], self.inputs['b'])}
+
+    def test_check_output(self):
+        self.check_output()
 
-    def test_normal(self):
-        self.check_grad(self.op, self.inputs, ["X", "b"], "Out")
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'b'], 'Out')
 
-    def test_ignore_b(self):
-        self.check_grad(self.op, self.inputs, ["X"], "Out", no_grad_set={"b"})
+    def test_check_grad_ignore_b(self):
+        self.check_grad(['X'], 'Out', no_grad_set=set('b'))
 
-    def test_ignore_x(self):
-        self.check_grad(self.op, self.inputs, ["b"], "Out", no_grad_set={"X"})
+    def test_check_grad_ignore_x(self):
+        self.check_grad(['b'], 'Out', no_grad_set=set('X'))
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_scale_and_identity_op.py b/python/paddle/v2/framework/tests/test_scale_and_identity_op.py
deleted file mode 100644
index 69b301c376..0000000000
--- a/python/paddle/v2/framework/tests/test_scale_and_identity_op.py
+++ /dev/null
@@ -1,43 +0,0 @@
-import unittest
-from op_test_util import OpTestMeta
-from gradient_checker import GradientChecker, create_op
-import numpy as np
-from paddle.v2.framework.op import Operator
-
-
-class IdentityTest(unittest.TestCase):
-    __metaclass__ = OpTestMeta
-
-    def setUp(self):
-        self.type = "identity"
-        self.inputs = {'X': np.random.random((32, 784)).astype("float32")}
-        self.outputs = {'Out': self.inputs['X']}
-
-
-class IdentityGradOpTest(GradientChecker):
-    def test_normal(self):
-        op = create_op("identity")
-        inputs = {"X": np.random.random((10, 10)).astype("float32")}
-        self.check_grad(op, inputs, set("X"), "Out")
-
-
-class ScaleTest(unittest.TestCase):
-    __metaclass__ = OpTestMeta
-
-    def setUp(self):
-        self.type = "scale"
-        self.inputs = {'X': np.random.random((32, 784)).astype("float32")}
-        self.attrs = {'scale': -2.3}
-        self.outputs = {'Out': self.inputs['X'] * self.attrs['scale']}
-
-
-class ScaleGradTest(GradientChecker):
-    def test_normal(self):
-        op = Operator("scale", X="X", Out="Out", scale=3.2)
-        self.check_grad(op,
-                        {"X": np.random.random((10, 10)).astype("float32")},
-                        set("X"), "Out")
-
-
-if __name__ == '__main__':
-    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_scale_op.py b/python/paddle/v2/framework/tests/test_scale_op.py
new file mode 100644
index 0000000000..2ea1e18547
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_scale_op.py
@@ -0,0 +1,21 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestScaleOp(OpTest):
+    def setUp(self):
+        self.op_type = "scale"
+        self.inputs = {'X': np.random.random((10, 10)).astype("float32")}
+        self.attrs = {'scale': -2.3}
+        self.outputs = {'Out': self.inputs['X'] * self.attrs['scale']}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Out')
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_scatter_op.py b/python/paddle/v2/framework/tests/test_scatter_op.py
index c1f9444889..33c73c5263 100644
--- a/python/paddle/v2/framework/tests/test_scatter_op.py
+++ b/python/paddle/v2/framework/tests/test_scatter_op.py
@@ -1,37 +1,24 @@
 import unittest
-from op_test_util import OpTestMeta
-from gradient_checker import GradientChecker, create_op
-import numpy
-import paddle.v2.framework.core as core
-from paddle.v2.framework.op import Operator
+import numpy as np
+from op_test import OpTest
 
 
-class TestScatterOp(unittest.TestCase):
-    __metaclass__ = OpTestMeta
-
+class TestScatterOp(OpTest):
     def setUp(self):
-        self.type = "scatter"
-        ref_np = numpy.ones((3, 3)).astype("float32")
-        index_np = numpy.array([1, 2]).astype("int32")
-        updates_np = numpy.random.random((2, 3)).astype("float32")
-        output_np = numpy.copy(ref_np)
+        self.op_type = "scatter"
+        ref_np = np.ones((3, 3)).astype("float32")
+        index_np = np.array([1, 2]).astype("int32")
+        updates_np = np.random.random((2, 3)).astype("float32")
+        output_np = np.copy(ref_np)
         output_np[index_np] += updates_np
         self.inputs = {'Ref': ref_np, 'Index': index_np, 'Updates': updates_np}
         self.outputs = {'Out': output_np}
 
+    def test_check_output(self):
+        self.check_output()
 
-class TestScatterGradOp(GradientChecker):
-    def test_scatter_grad(self):
-        op = create_op("scatter")
-        # test data setup
-        ref_np = numpy.ones((3, 10)).astype("float32")
-        index_np = numpy.array([1, 2]).astype("int32")
-        updates_np = numpy.random.random((2, 10)).astype("float32")
-        output_np = numpy.copy(ref_np)
-        output_np[index_np] += updates_np
-        inputs = {'Ref': ref_np, 'Index': index_np, 'Updates': updates_np}
-        self.check_grad(
-            op, inputs, set(["Updates", "Ref"]), "Out", in_place=True)
+    def test_check_grad(self):
+        self.check_grad(['Updates', 'Ref'], 'Out', in_place=True)
 
 
 if __name__ == "__main__":
diff --git a/python/paddle/v2/framework/tests/test_seq_pool.py b/python/paddle/v2/framework/tests/test_seq_pool.py
new file mode 100644
index 0000000000..cf864936af
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_seq_pool.py
@@ -0,0 +1,51 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestSeqAvgPool1D(OpTest):
+    def setUp(self):
+        self.op_type = 'sequence_avg_pool'
+        # one level, batch size is 4
+        x = np.random.uniform(0.1, 1, [11, 23]).astype('float32')
+        lod = [[0, 4, 5, 8, 11]]
+
+        out = np.zeros((4, 23)).astype('float32')
+        for i in range(4):
+            sub_x = x[lod[0][i]:lod[0][i + 1], :]
+            out[i] = sub_x.mean(axis=0)
+
+        self.inputs = {'X': (x, lod)}
+        self.outputs = {'Out': out}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(["X"], "Out")
+
+
+class TestSeqAvgPool2D(OpTest):
+    def setUp(self):
+        self.op_type = 'sequence_avg_pool'
+        # one level, batch size is 4
+        x = np.random.uniform(0.1, 1, [13, 3, 17]).astype('float32')
+        lod = [[0, 4, 5, 8, 13]]
+
+        out = np.zeros((4, 3, 17)).astype('float32')
+        for i in range(4):
+            sub_x = np.reshape(x[lod[0][i]:lod[0][i + 1], :], (-1, 3 * 17))
+            out[i] = np.reshape(sub_x.mean(axis=0), (3, 17))
+
+        self.inputs = {'X': (x, lod)}
+        self.outputs = {'Out': out}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(["X"], "Out")
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_sgd_op.py b/python/paddle/v2/framework/tests/test_sgd_op.py
index e5f9ef865e..64e54d1500 100644
--- a/python/paddle/v2/framework/tests/test_sgd_op.py
+++ b/python/paddle/v2/framework/tests/test_sgd_op.py
@@ -1,21 +1,22 @@
 import unittest
-import numpy
-from op_test_util import OpTestMeta
+import numpy as np
+from op_test import OpTest
 
 
-class TestSGD(unittest.TestCase):
-    __metaclass__ = OpTestMeta
-
+class TestSGDOp(OpTest):
     def setUp(self):
-        self.type = "sgd"
-        w = numpy.random.random((102, 105)).astype("float32")
-        g = numpy.random.random((102, 105)).astype("float32")
+        self.op_type = "sgd"
+        w = np.random.random((102, 105)).astype("float32")
+        g = np.random.random((102, 105)).astype("float32")
         lr = 0.1
 
         self.inputs = {'param': w, 'grad': g}
         self.attrs = {'learning_rate': lr}
         self.outputs = {'param_out': w - lr * g}
 
+    def test_check_output(self):
+        self.check_output()
+
 
 if __name__ == "__main__":
     unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_sigmoid_op.py b/python/paddle/v2/framework/tests/test_sigmoid_op.py
index 273c2e5ab1..d65d887db4 100644
--- a/python/paddle/v2/framework/tests/test_sigmoid_op.py
+++ b/python/paddle/v2/framework/tests/test_sigmoid_op.py
@@ -1,27 +1,21 @@
 import unittest
 import numpy as np
-from op_test_util import OpTestMeta
-from gradient_checker import GradientChecker, create_op
+from op_test import OpTest
 
 
-class TestSigmoidOp(unittest.TestCase):
-    __metaclass__ = OpTestMeta
-
+class TestSigmoidOp(OpTest):
     def setUp(self):
-        self.type = "sigmoid"
-        self.inputs = {'X': np.random.random((15, 31)).astype("float32")}
+        self.op_type = "sigmoid"
+        self.inputs = {
+            'X': np.random.uniform(0.1, 1, [11, 17]).astype("float32")
+        }
         self.outputs = {'Y': 1 / (1 + np.exp(-self.inputs['X']))}
 
+    def test_check_output(self):
+        self.check_output()
 
-class TestSigmoidGradOp(GradientChecker):
-    def test_grad(self):
-        op = create_op("sigmoid")
-        inputs = {"X": np.random.uniform(0.1, 1, [11, 17]).astype("float32")}
-        # compare gpu and cpu results for backward op.
-        # this test will be skiped if only compiling CPU version.
-        self.compare_grad(op, inputs)
-        # check gradients 
-        self.check_grad(op, inputs, set("X"), "Y", max_relative_error=0.007)
+    def test_check_grad(self):
+        self.check_grad(["X"], "Y", max_relative_error=0.007)
 
 
 if __name__ == '__main__':
diff --git a/python/paddle/v2/framework/tests/test_softmax_op.py b/python/paddle/v2/framework/tests/test_softmax_op.py
index 0d590fa706..1b948f252f 100644
--- a/python/paddle/v2/framework/tests/test_softmax_op.py
+++ b/python/paddle/v2/framework/tests/test_softmax_op.py
@@ -1,9 +1,6 @@
 import unittest
-
 import numpy as np
-
-from gradient_checker import GradientChecker, create_op
-from op_test_util import OpTestMeta
+from op_test import OpTest
 
 
 def stable_softmax(x):
@@ -13,26 +10,21 @@ def stable_softmax(x):
     return exps / np.sum(exps)
 
 
-class TestSoftmaxOp(unittest.TestCase):
-    __metaclass__ = OpTestMeta
-
+class TestSoftmaxOp(OpTest):
     def setUp(self):
-        self.type = "softmax"
-        self.inputs = {"X": np.random.random((10, 10)).astype("float32")}
+        self.op_type = "softmax"
+        self.inputs = {
+            'X': np.random.uniform(0.1, 1, [10, 10]).astype("float32")
+        }
         self.outputs = {
-            "Y": np.apply_along_axis(stable_softmax, 1, self.inputs["X"])
+            'Y': np.apply_along_axis(stable_softmax, 1, self.inputs['X'])
         }
 
+    def test_check_output(self):
+        self.check_output()
 
-class TestSoftmaxGradOp(GradientChecker):
-    def setUp(self):
-        self.op = create_op("softmax")
-        self.inputs = {
-            "X": np.random.uniform(0.1, 1, [10, 10]).astype("float32")
-        }
-
-    def test_softmax_grad(self):
-        self.check_grad(self.op, self.inputs, ["X"], "Y")
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Y')
 
 
 if __name__ == "__main__":
diff --git a/python/paddle/v2/framework/tests/test_split_op.py b/python/paddle/v2/framework/tests/test_split_op.py
new file mode 100644
index 0000000000..b4420db9d7
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_split_op.py
@@ -0,0 +1,26 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestSplitOp(OpTest):
+    def setUp(self):
+        self.op_type = "split"
+        axis = 0
+        num = 2
+        x = np.random.random((4, 2)).astype('float32')
+        out = np.split(x, num, axis)
+        self.inputs = {'X': x}
+        self.attrs = {'axis': axis, 'num': num}
+        self.outputs = {'Out': [('out%d' % i, out[i]) \
+            for i in xrange(len(out))]}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], ['out0', 'out1'])
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_squared_l2_distance_op.py b/python/paddle/v2/framework/tests/test_squared_l2_distance_op.py
index 2bcdf37df4..dc6ebf5d30 100644
--- a/python/paddle/v2/framework/tests/test_squared_l2_distance_op.py
+++ b/python/paddle/v2/framework/tests/test_squared_l2_distance_op.py
@@ -1,17 +1,14 @@
 import unittest
-from op_test_util import OpTestMeta
-from gradient_checker import GradientChecker, create_op
 import numpy as np
+from op_test import OpTest
 
 
-class TestSquaredL2DistanceOp_f0(unittest.TestCase):
-    __metaclass__ = OpTestMeta
-
+class TestSquaredL2DistanceOp_f0(OpTest):
     def setUp(self):
-        self.type = 'squared_l2_distance'
+        self.op_type = "squared_l2_distance"
         self.inputs = {
-            'X': np.random.uniform(0.1, 1., (32, 64)).astype('float32'),
-            'Y': np.random.uniform(0.1, 1., (32, 64)).astype('float32')
+            'X': np.random.uniform(0.1, 0.6, (2, 3)).astype("float32"),
+            'Y': np.random.uniform(0.1, 0.6, (2, 3)).astype("float32")
         }
         sub_res = self.inputs['X'] - self.inputs['Y']
         output = sub_res * sub_res
@@ -20,15 +17,19 @@ class TestSquaredL2DistanceOp_f0(unittest.TestCase):
             'Out': np.expand_dims(output.sum(1), 1)
         }
 
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X', 'Y'], 'Out')
 
-class TestSquaredL2DistanceOp_f1(unittest.TestCase):
-    __metaclass__ = OpTestMeta
 
+class TestSquaredL2DistanceOp_f1(OpTest):
     def setUp(self):
-        self.type = 'squared_l2_distance'
+        self.op_type = "squared_l2_distance"
         self.inputs = {
-            'X': np.random.uniform(0.1, 1., (32, 64)).astype('float32'),
-            'Y': np.random.uniform(0.1, 1., (1, 64)).astype('float32')
+            'X': np.random.uniform(0.1, 0.6, (2, 3)).astype("float32"),
+            'Y': np.random.uniform(0.1, 0.6, (1, 3)).astype("float32")
         }
         sub_res = self.inputs['X'] - self.inputs['Y']
         output = sub_res * sub_res
@@ -37,53 +38,34 @@ class TestSquaredL2DistanceOp_f1(unittest.TestCase):
             'Out': np.expand_dims(output.sum(1), 1)
         }
 
+    def test_check_output(self):
+        self.check_output()
 
-class TestSquaredL2DistanceOp_f2(unittest.TestCase):
-    __metaclass__ = OpTestMeta
+    def test_check_grad(self):
+        self.check_grad(['X', 'Y'], 'Out')
 
+
+class TestSquaredL2DistanceOp_f2(OpTest):
     def setUp(self):
-        self.type = 'squared_l2_distance'
+        self.op_type = "squared_l2_distance"
         self.inputs = {
-            'X': np.random.uniform(0.1, 1., (32, 64, 128)).astype('float32'),
-            'Y': np.random.uniform(0.1, 1., (1, 64, 128)).astype('float32')
+            'X': np.random.uniform(0.1, 0.6, (2, 3, 4)).astype("float32"),
+            'Y': np.random.uniform(0.1, 0.6, (1, 3, 4)).astype("float32")
         }
         sub_res = self.inputs['X'] - self.inputs['Y']
-        sub_res = sub_res.reshape((32, 64 * 128))
+        sub_res = sub_res.reshape((2, 3 * 4))
         output = sub_res * sub_res
         self.outputs = {
             'sub_result': sub_res,
             'Out': np.expand_dims(output.sum(1), 1)
         }
 
+    def test_check_output(self):
+        self.check_output()
 
-class TestSquaredL2DistanceGradOp(GradientChecker):
-    def test_squared_l2_distance_b0(self):
-        op = create_op("squared_l2_distance")
-        inputs = {
-            'X': np.random.uniform(0.1, .6, (2, 3)).astype('float32'),
-            'Y': np.random.uniform(0.1, .6, (2, 3)).astype('float32')
-        }
-        self.compare_grad(op, inputs)
-        self.check_grad(op, inputs, set(["X", "Y"]), "Out")
-
-    def test_squared_l2_distance_b1(self):
-        op = create_op("squared_l2_distance")
-        inputs = {
-            'X': np.random.uniform(0.1, .6, (2, 3)).astype('float32'),
-            'Y': np.random.uniform(0.1, .6, (1, 3)).astype('float32')
-        }
-        self.compare_grad(op, inputs)
-        self.check_grad(op, inputs, set(["X", "Y"]), "Out")
-
-    def test_squared_l2_distance_b2(self):
-        op = create_op("squared_l2_distance")
-        inputs = {
-            'X': np.random.uniform(0.1, .6, (2, 3, 4)).astype('float32'),
-            'Y': np.random.uniform(0.1, .6, (1, 3, 4)).astype('float32')
-        }
-        self.compare_grad(op, inputs)
-        self.check_grad(op, inputs, set(["X", "Y"]), "Out")
+    def test_check_grad(self):
+        self.check_grad(['X', 'Y'], 'Out')
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_sum_op.py b/python/paddle/v2/framework/tests/test_sum_op.py
new file mode 100644
index 0000000000..60254291e2
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_sum_op.py
@@ -0,0 +1,24 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestSumOp(OpTest):
+    def setUp(self):
+        self.op_type = "sum"
+        x0 = np.random.random((3, 4)).astype('float32')
+        x1 = np.random.random((3, 4)).astype('float32')
+        x2 = np.random.random((3, 4)).astype('float32')
+        self.inputs = {"X": [("x0", x0), ("x1", x1), ("x2", x2)]}
+        y = x0 + x1 + x2
+        self.outputs = {'Out': y}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['x0'], 'Out')
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_tensor.py b/python/paddle/v2/framework/tests/test_tensor.py
index 1af39818a3..8cd93b35d7 100644
--- a/python/paddle/v2/framework/tests/test_tensor.py
+++ b/python/paddle/v2/framework/tests/test_tensor.py
@@ -3,7 +3,7 @@ import unittest
 import numpy
 
 
-class TestScope(unittest.TestCase):
+class TestTensor(unittest.TestCase):
     def test_int_tensor(self):
         scope = core.Scope()
         var = scope.new_var("test_tensor")
@@ -20,8 +20,8 @@ class TestScope(unittest.TestCase):
         tensor.set(tensor_array, place)
 
         tensor_array_2 = numpy.array(tensor)
-        self.assertEqual(1.0, tensor_array_2[3, 9])
-        self.assertEqual(2.0, tensor_array_2[19, 11])
+        self.assertEqual(1, tensor_array_2[3, 9])
+        self.assertEqual(2, tensor_array_2[19, 11])
 
     def test_float_tensor(self):
         scope = core.Scope()
@@ -43,6 +43,71 @@ class TestScope(unittest.TestCase):
         self.assertAlmostEqual(1.0, tensor_array_2[3, 9])
         self.assertAlmostEqual(2.0, tensor_array_2[19, 11])
 
+    def test_int_lod_tensor(self):
+        place = core.CPUPlace()
+        scope = core.Scope()
+        var_lod = scope.new_var("test_lod_tensor")
+        lod_tensor = var_lod.get_tensor()
+
+        lod_tensor.set_dims([4, 4, 6])
+        lod_tensor.alloc_int(place)
+        array = numpy.array(lod_tensor)
+        array[0, 0, 0] = 3
+        array[3, 3, 5] = 10
+        lod_tensor.set(array, place)
+        lod_tensor.set_lod([[0, 2, 4]])
+
+        lod_v = numpy.array(lod_tensor)
+        self.assertTrue(numpy.alltrue(array == lod_v))
+
+        lod = lod_tensor.lod()
+        self.assertEqual(0, lod[0][0])
+        self.assertEqual(2, lod[0][1])
+        self.assertEqual(4, lod[0][2])
+
+    def test_float_lod_tensor(self):
+        place = core.CPUPlace()
+        scope = core.Scope()
+        var_lod = scope.new_var("test_lod_tensor")
+
+        lod_tensor = var_lod.get_tensor()
+        lod_tensor.set_dims([5, 2, 3, 4])
+        lod_tensor.alloc_float(place)
+
+        tensor_array = numpy.array(lod_tensor)
+        self.assertEqual((5, 2, 3, 4), tensor_array.shape)
+        tensor_array[0, 0, 0, 0] = 1.0
+        tensor_array[0, 0, 0, 1] = 2.0
+        lod_tensor.set(tensor_array, place)
+
+        lod_v = numpy.array(lod_tensor)
+        self.assertAlmostEqual(1.0, lod_v[0, 0, 0, 0])
+        self.assertAlmostEqual(2.0, lod_v[0, 0, 0, 1])
+        self.assertEqual(len(lod_tensor.lod()), 0)
+
+        lod_py = [[0, 2, 5], [0, 2, 4, 5]]
+        lod_tensor.set_lod(lod_py)
+        lod = lod_tensor.lod()
+        self.assertListEqual(lod_py, lod)
+
+    def test_lod_tensor_init(self):
+        scope = core.Scope()
+        place = core.CPUPlace()
+        lod_py = [[0, 2, 5], [0, 2, 4, 5]]
+        lod_tensor = core.LoDTensor(lod_py)
+
+        lod_tensor.set_dims([5, 2, 3, 4])
+        lod_tensor.alloc_float(place)
+        tensor_array = numpy.array(lod_tensor)
+        tensor_array[0, 0, 0, 0] = 1.0
+        tensor_array[0, 0, 0, 1] = 2.0
+        lod_tensor.set(tensor_array, place)
+
+        lod_v = numpy.array(lod_tensor)
+        self.assertAlmostEqual(1.0, lod_v[0, 0, 0, 0])
+        self.assertAlmostEqual(2.0, lod_v[0, 0, 0, 1])
+        self.assertListEqual(lod_py, lod_tensor.lod())
+
 
 if __name__ == '__main__':
     unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_top_k_op.py b/python/paddle/v2/framework/tests/test_top_k_op.py
index e841d96d26..694f37d612 100644
--- a/python/paddle/v2/framework/tests/test_top_k_op.py
+++ b/python/paddle/v2/framework/tests/test_top_k_op.py
@@ -1,14 +1,11 @@
 import unittest
 import numpy as np
-from gradient_checker import GradientChecker, create_op
-from op_test_util import OpTestMeta
+from op_test import OpTest
 
 
-class TestTopkOp(unittest.TestCase):
-    __metaclass__ = OpTestMeta
-
+class TestTopkOp(OpTest):
     def setUp(self):
-        self.type = "top_k"
+        self.op_type = "top_k"
         k = 1
         input = np.random.random((32, 84)).astype("float32")
         output = np.ndarray((32, k))
@@ -24,12 +21,13 @@ class TestTopkOp(unittest.TestCase):
 
         self.outputs = {'Out': output, 'Indices': indices}
 
+    def test_check_output(self):
+        self.check_output()
 
-class TestTopkOp3d(unittest.TestCase):
-    __metaclass__ = OpTestMeta
 
+class TestTopkOp3d(OpTest):
     def setUp(self):
-        self.type = "top_k"
+        self.op_type = "top_k"
         k = 1
         input = np.random.random((32, 2, 84)).astype("float32")
         input_flat_2d = input.reshape(64, 84)
@@ -47,6 +45,9 @@ class TestTopkOp3d(unittest.TestCase):
 
         self.outputs = {'Out': output, 'Indices': indices}
 
+    def test_check_output(self):
+        self.check_output()
+
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_uniform_random_op.py b/python/paddle/v2/framework/tests/test_uniform_random_op.py
index c3d2bb44da..9e8898fb59 100644
--- a/python/paddle/v2/framework/tests/test_uniform_random_op.py
+++ b/python/paddle/v2/framework/tests/test_uniform_random_op.py
@@ -4,7 +4,7 @@ import paddle.v2.framework.core as core
 import numpy
 
 
-class UniformRandomTest(unittest.TestCase):
+class TestUniformRandomOp(unittest.TestCase):
     def test_uniform_random_cpu(self):
         self.uniform_random_test(place=core.CPUPlace())
 
@@ -14,11 +14,11 @@ class UniformRandomTest(unittest.TestCase):
 
     def uniform_random_test(self, place):
         scope = core.Scope()
-        scope.new_var("X").get_tensor()
+        scope.new_var('X').get_tensor()
 
         op = Operator(
             "uniform_random",
-            Out="X",
+            Out='X',
             dims=[1000, 784],
             min=-5.0,
             max=10.0,
@@ -27,9 +27,9 @@ class UniformRandomTest(unittest.TestCase):
         op.infer_shape(scope)
         ctx = core.DeviceContext.create(place)
         op.run(scope, ctx)
-        tensor = numpy.array(scope.find_var("X").get_tensor())
+        tensor = numpy.array(scope.find_var('X').get_tensor())
         self.assertAlmostEqual(tensor.mean(), 2.5, delta=0.1)
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     unittest.main()
diff --git a/python/paddle/v2/inference.py b/python/paddle/v2/inference.py
index 8acea6155c..e80456d9bb 100644
--- a/python/paddle/v2/inference.py
+++ b/python/paddle/v2/inference.py
@@ -2,6 +2,7 @@ import numpy
 import collections
 import topology
 import minibatch
+import cPickle
 
 __all__ = ['infer', 'Inference']
 
@@ -25,11 +26,23 @@ class Inference(object):
     :type parameters: paddle.v2.parameters.Parameters
     """
 
-    def __init__(self, output_layer, parameters):
+    def __init__(self, parameters, output_layer=None, fileobj=None):
         import py_paddle.swig_paddle as api
-        topo = topology.Topology(output_layer)
-        gm = api.GradientMachine.createFromConfigProto(
-            topo.proto(), api.CREATE_MODE_TESTING, [api.PARAMETER_VALUE])
+
+        if output_layer is not None:
+            topo = topology.Topology(output_layer)
+            gm = api.GradientMachine.createFromConfigProto(
+                topo.proto(), api.CREATE_MODE_TESTING, [api.PARAMETER_VALUE])
+            self.__data_types__ = topo.data_type()
+        elif fileobj is not None:
+            tmp = cPickle.load(fileobj)
+            gm = api.GradientMachine.createByConfigProtoStr(
+                tmp['protobin'], api.CREATE_MODE_TESTING,
+                [api.PARAMETER_VALUE])
+            self.__data_types__ = tmp['data_type']
+        else:
+            raise ValueError("Either output_layer or fileobj must be set")
+
         for param in gm.getParameters():
             val = param.getBuf(api.PARAMETER_VALUE)
             name = param.getName()
@@ -43,7 +56,6 @@ class Inference(object):
             # called here, but it's better to call this function in one place.
             param.setValueUpdated()
         self.__gradient_machine__ = gm
-        self.__data_types__ = topo.data_type()
 
     def iter_infer(self, input, feeding=None):
         from data_feeder import DataFeeder
diff --git a/python/paddle/v2/topology.py b/python/paddle/v2/topology.py
index a20e878d08..2db66be250 100644
--- a/python/paddle/v2/topology.py
+++ b/python/paddle/v2/topology.py
@@ -18,6 +18,7 @@ from paddle.proto.ModelConfig_pb2 import ModelConfig
 import paddle.trainer_config_helpers as conf_helps
 import layer as v2_layer
 import config_base
+import cPickle
 
 __all__ = ['Topology']
 
@@ -100,6 +101,14 @@ class Topology(object):
                 return layer
         return None
 
+    def serialize_for_inference(self, stream):
+        protobin = self.proto().SerializeToString()
+        data_type = self.data_type()
+        cPickle.dump({
+            'protobin': protobin,
+            'data_type': data_type
+        }, stream, cPickle.HIGHEST_PROTOCOL)
+
 
 def __check_layer_type__(layer):
     if not isinstance(layer, config_base.Layer):
diff --git a/python/paddle/v2/trainer.py b/python/paddle/v2/trainer.py
index 0654a30104..ca95ef13bd 100644
--- a/python/paddle/v2/trainer.py
+++ b/python/paddle/v2/trainer.py
@@ -174,13 +174,18 @@ class SGD(object):
                         pass_id=pass_id,
                         batch_id=batch_id,
                         cost=cost,
-                        evaluator=batch_evaluator))
+                        evaluator=batch_evaluator,
+                        gm=self.__gradient_machine__))
                 self.__parameter_updater__.finishBatch(cost)
                 batch_evaluator.finish()
 
             self.__parameter_updater__.finishPass()
             pass_evaluator.finish()
-            event_handler(v2_event.EndPass(pass_id, evaluator=pass_evaluator))
+            event_handler(
+                v2_event.EndPass(
+                    pass_id,
+                    evaluator=pass_evaluator,
+                    gm=self.__gradient_machine__))
         self.__gradient_machine__.finish()
 
     def test(self, reader, feeding=None):