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

.gitignore

@@ -22,7 +22,9 @@ cmake-build-*
 
 # generated while compiling
 python/paddle/v2/framework/core.so
+paddle/pybind/pybind.h
 CMakeFiles
 cmake_install.cmake
 paddle/.timestamp
 python/paddlepaddle.egg-info/
+paddle/pybind/pybind.h

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

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"]

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
 #

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()

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()

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}")

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()

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()

doc/design/block.md

+# 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);
+}
+```

doc/design/if_else_op.md

-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
 

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

+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
+}

doc/design/ops/images/rnn.dot

+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]
+
+}

doc/design/ops/images/rnn_2level_data.dot

+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
+
+}

doc/design/ops/rnn.md

+# RNNOp design
+
+This document is about an RNN operator which requires that instances in a mini-batch have the same length.  We will have a more flexible RNN operator.
+
+## 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>

doc/design/refactorization.md

+# Design Doc: Refactorization Overview
+
+The goal of refactorizaiton include:
+
+1. Make it easy for external contributors to write new elementory computaiton operations.
+1. Make the codebase clean and readable.
+1. Introduce a new design of computation representation -- a computation graph of operators and variables.
+1. The graph representation helps implementing auto-scalable and auto fault recoverable distributed computing.
+
+## Computation Graphs
+
+1. PaddlePaddle represent the computation, training and inference of DL models, by computation graphs.
+
+  1. Please dig into [computation graphs](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/graph.md) for a solid example.
+
+1. Users write Python programs to describe the graphs and run it (locally or remotely).
+
+1. A graph is composed of *variabels* and *operators*.
+
+1. The description of graphs must be able to be serialized/deserialized, so it
+
+   1. could to be sent to the cloud for distributed execution, and
+   1. be sent to clients for mobile or enterprise deployment.
+
+1. The Python program do
+
+   1. *compilation*: runs a Python program to generate a protobuf message representation of the graph and send it to
+      1. the C++ library `libpaddle.so` for local execution,
+      1. the master process of a distributed training job for training, or
+      1. the server process of a Kubernetes serving job for distributed serving.
+   1. *execution*: according to the protobuf message, constructs instances of class `Variable` and `OperatorBase`, and run them.
+
+## Description and Realization
+
+At compile time, the Python program generates protobuf message representation of the graph, or the description of the graph.
+
+At runtime, the C++ program realizes the graph and run it.
+
+| | Representation (protobuf messages) | Realization (C++ class objects) |
+|---|---|---|
+|Data|[VarDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto#L107)|[Variable](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/variable.h#L24)|
+|Operation|[OpDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto#L35)|[Operator](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h#L64)|
+|Block|BlockDesc|Block|
+
+The word *graph* is exchangable with *block* in this document.  A graph represent computation steps and local variables as a C++/Java program block, or a pair of { and }.
+
+## Compilation and Execution
+
+1. Run an applicaton Python program to describe the graph.  In particular,
+
+   1. create VarDesc to represent local/intermediate variables,
+   1. create operators and set attributes,
+   1. validate attribute values,
+   1. inference the type and the shape of variables,
+   1. plan for memory-reuse for variables,
+   1. generate backward and optimization part of the Graph.
+   1. possiblly split the graph for distributed training.
+
+1. The invocation of `train` or `infer` in the application Python program:
+
+   1. create a new Scope instance in the [scope hierarchy](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/scope.md) for each run of a block,
+      1. realize local variables defined in the BlockDesc message in the new scope,
+      1. a scope is similar to the stack frame in programming languages,
+
+   1. create an instance of class `Block`, in which,
+      1. realize operators in the BlockDesc message,
+
+   1. run the Block by calling
+      1. `Block::Eval(vector<Variable>* targets)` for forward and backward computations, or
+      1. `Block::Eval(vector<Operator>* targets)` for optimization.
+
+
+## Intermediate Representation (IR)
+
+```text
+Compile Time -> IR -> Runtime
+```
+
+### Benefit
+
+- Optimization
+  ```text
+  Compile Time -> IR -> Optimized IR -> Runtime
+  ```
+- Send automatically partitioned IR to different nodes.
+  - Automatic data parallel
+    ```text
+    Compile Time
+    |-> Single GPU IR
+        |-> [trainer-IR-0, trainer-IR-1, pserver-IR]
+            |-> Node-0 (runs trainer-IR-0)
+            |-> Node-1 (runs trainer-IR-1)
+            |-> Node-2 (runs pserver-IR)
+    ```
+  - Automatic model parallel (planned for future)
+
+---
+
+# Operator/OpWithKernel/OpKernel
+
+![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/49caf1fb70820fb4a6c217634317c9306f361f36/op_op_with_kern_class_diagram.dot)
+
+---
+
+# Operator
+![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/dd598e8f1976f5759f58af5e5ef94738a6b2e661/op.dot)
+
+* `Operator` is the fundamental building block as the user interface.
+    * Operator stores input/output variable name, and attributes.
+    * The `InferShape` interface is used to infer output variable shapes by its input shapes.
+    * Use `Run` to compute `input variables` to `output variables`.
+
+---
+
+# OpWithKernel/Kernel
+
+![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/9d7f4eba185cf41c8e2fbfb40ae21890dbddcd39/op_with_kernel.dot)
+
+* `OpWithKernel` inherits `Operator`.
+* `OpWithKernel` contains a Kernel map.
+    * `OpWithKernel::Run` get device's kernel, and invoke `OpKernel::Compute`.
+    * `OpKernelKey` is the map key. Only device place now, but may be data type later.
+
+---
+
+# Why separate Kernel and Operator
+
+* Separate GPU and CPU code.
+    * Make Paddle can run without GPU.
+* Make one operator (which is user interface) can contain many implementations.
+    * Same mul op, different FP16, FP32 Kernel. different MKL, eigen kernel.
+---
+
+# Libraries for Kernel development
+
+* `Eigen::Tensor` contains basic math and element-wise functions.
+    * Note that `Eigen::Tensor` has broadcast implementation.
+    * Limit number of `tensor.device(dev) = ` in your code.
+* `thrust::tranform` and `std::transform`.
+    * `thrust` has the same API as C++ standard library. Using `transform` can quickly implement a customized elementwise kernel.
+    * `thrust` has more complex API, like `scan`, `reduce`, `reduce_by_key`.
+* Hand-writing `GPUKernel` and `CPU` code
+    * Do not write `.h`. CPU Kernel should be in `.cc`. CPU kernel should be in `.cu`. (`GCC` cannot compile GPU code.)
+---
+# Operator Register
+
+## Why register is necessary?
+We need a method to build mappings between Op type names and Op classes.
+
+## How to do the register?
+
+Maintain a map, whose key is the type name and value is corresponding Op constructor.
+
+---
+# The Registry Map
+
+### `OpInfoMap`
+
+`op_type(string)` -> `OpInfo`
+
+`OpInfo`:
+
+- **`creator`**: The Op constructor.
+- **`grad_op_type`**: The type of the gradient Op.
+- **`proto`**: The Op's Protobuf, including inputs, outputs and required attributes.
+- **`checker`**: Used to check attributes.
+
+---
+# Related Concepts
+
+### Op_Maker
+It's constructor takes `proto` and `checker`. They are compeleted during Op_Maker's construction. ([ScaleOpMaker](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37))
+
+### Register Macros
+```cpp
+REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type, grad_op_class)
+REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class)
+```
+
+### `USE` Macros
+make sure the registration process is executed and linked.
+
+---
+# Register Process
+1. Write Op class, as well as its gradient Op class if there is.
+2. Write Op maker class. In the constructor, describe its inputs, outputs, and attributes.
+3. Invoke macro `REGISTER_OP`. The macro will
+	1. call maker class to complete `proto` and `checker`
+	2. with the completed `proto` and `checker`, build a new key-value pair in the `OpInfoMap`
+
+4. Invoke `USE` macro in where the Op is used to make sure it is linked.
+
+---
+# Backward Module (1/2)
+### Create Backward Operator
+- Mapping from forwarding Op to backward Op
+![backward](https://gist.githubusercontent.com/dzhwinter/a6fbd4623ee76c459f7f94591fd1abf0/raw/61026ab6e518e66bde66a889bc42557a1fccff33/backward.png)
+
+---
+# Backward Module (2/2)
+### Build Backward Network
+- **Input** graph of forwarding operators
+- **Output** graph of backward operators
+- **corner case in construction**
+	- shared variable => insert `Add` operator
+	- no gradient => insert `fill_zero_grad` operator
+	- recursive netOp => call `Backward` recursively
+	- RNN Op => recursively call `Backward` on stepnet
+
+
+---
+# Scope, Variable, Tensor
+
+* `Tensor` is an n-dimension array with type.
+	* Only dims and data pointers are stored in `Tensor`.
+	* All operators on `Tensor` is written in `Operator` or global functions.
+	* variable length Tensor design [LoDTensor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/lod_tensor.md)
+* `Variable` is the inputs and outputs of an operator. Not just `Tensor`.
+	* step_scopes in RNN is a variable and not a tensor.
+* `Scope` is where variables store at.
+	* map<string/*var name */, Variable>
+	* `Scope` has a hierarchical structure. The local scope can get variable from its parent scope.
+
+---
+# Block (in design)
+## the difference with original RNNOp
+- as an operator is more intuitive than `RNNOp`,
+- offers new interface `Eval(targets)` to deduce the minimal block to `Run`,
+- fits the compile-time/ runtime separation design.
+  - during the compilation, `SymbolTable` stores `VarDesc`s and `OpDesc`s and serialize to a `BlockDesc`
+  - when graph executes, a Block with `BlockDesc` passed in creates `Op` and `Var` then `Run`
+
+---
+# Milestone
+- take Paddle/books as the main line, the requirement of the models motivates framework refactoring,
+- model migration
+  - framework development gives **priority support** to model migration, for example,
+    - the MNIST demo needs a Python interface,
+    - the RNN models require the framework to support `LoDTensor`.
+  - determine some timelines,
+  - heavily-relied Ops need to be migrated first,
+  - different models can be migrated parallelly.
+- improve the framework at the same time
+- accept imperfection, concentrated on solving the specific problem at the right price.
+
+---
+# Control the migration quality
+- compare the performance of migrated models with old ones.
+- follow google C style
+- build the automatic workflow of generating Python/C++ documentations
+  - the documentation of layers and ops should be written inside the code
+  - take the documentation quality into account when doing PR
+  - preview the documentations, read and improve them from users' perspective

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
-
-   `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
 
@@ -367,3 +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。

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)

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;
 

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);
   }
 };
 

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;
         }

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";
   }

paddle/framework/lod_tensor.h

@@ -51,18 +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.
@@ -104,7 +101,6 @@ class LoDTensor {
 
  private:
   LoD lod_;
-  Tensor* tensor_;  // not owned
 };
 }  // namespace framework
 }  // namespace paddle

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

paddle/framework/lod_tensor_test.cu

@@ -26,18 +26,16 @@ __global__ void test(size_t* a, int size) {
 }
 
 TEST(LoDTensor, LoDInGPU) {
-  paddle::framework::Tensor tensor;
   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});
 
-  tensor.Resize({14, 16});
-  tensor.mutable_data<float>(place);
+  lod_tensor.Resize({14, 16});
+  lod_tensor.mutable_data<float>(place);
 
   lod_tensor.set_lod(src_lod);
-  lod_tensor.set_tensor(&tensor);
   CHECK_EQ(lod_tensor.lod_element(0, 2), 4);
   CHECK_EQ(lod_tensor.lod_element(0, 4), 8);
 

paddle/framework/operator.cc

@@ -22,14 +22,14 @@ namespace framework {
 template <>
 Eigen::DefaultDevice& ExecutionContext::GetEigenDevice<
     platform::CPUPlace, Eigen::DefaultDevice>() const {
-  return *device_context_->get_eigen_device<Eigen::DefaultDevice>();
+  return *device_context_.get_eigen_device<Eigen::DefaultDevice>();
 }
 
 #ifndef PADDLE_ONLY_CPU
 template <>
 Eigen::GpuDevice&
 ExecutionContext::GetEigenDevice<platform::GPUPlace, Eigen::GpuDevice>() const {
-  return *device_context_->get_eigen_device<Eigen::GpuDevice>();
+  return *device_context_.get_eigen_device<Eigen::GpuDevice>();
 }
 #endif
 
@@ -186,6 +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();

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"
@@ -326,11 +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;
 
@@ -349,7 +366,7 @@ struct EigenDeviceConverter<platform::GPUPlace> {
 class ExecutionContext : public InferShapeContext {
  public:
   ExecutionContext(const OperatorBase& op, const Scope& scope,
-                   const platform::DeviceContext* device_context)
+                   const platform::DeviceContext& device_context)
       : InferShapeContext(op, scope), device_context_(device_context) {}
 
   template <typename PlaceType,
@@ -357,15 +374,44 @@ class ExecutionContext : public InferShapeContext {
                 typename EigenDeviceConverter<PlaceType>::EigenDeviceType>
   DeviceType& GetEigenDevice() const;
 
-  platform::Place GetPlace() const { return device_context_->GetPlace(); }
+  platform::Place GetPlace() const { return device_context_.GetPlace(); }
 
-  const platform::DeviceContext* device_context() const {
+  const platform::DeviceContext& device_context() const {
     return device_context_;
   }
 
-  const platform::DeviceContext* 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;
+  }
+
+ private:
+  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:
   /**
@@ -416,7 +462,7 @@ class OperatorWithKernel : public OperatorBase {
   void Run(const Scope& scope,
            const platform::DeviceContext& dev_ctx) const final {
     auto& opKernel = AllOpKernels().at(type_).at(OpKernelKey(dev_ctx));
-    opKernel->Compute(ExecutionContext(*this, scope, &dev_ctx));
+    opKernel->Compute(ExecutionContext(*this, scope, dev_ctx));
   }
 
   static std::unordered_map<std::string /* op_type */, OpKernelMap>&

paddle/framework/tensor_impl.h

@@ -22,7 +22,7 @@ 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(), numel() * sizeof(T) + offset_,
       "Tensor's dims_ is out of bound. Call Tensor::mutable_data "

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
+}

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;

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;

paddle/gserver/layers/DetectionOutputLayer.cpp

@@ -143,7 +143,7 @@ void DetectionOutputLayer::forward(PassType passType) {
     resetOutput(numKept, 7);
   } else {
     MatrixPtr outV = getOutputValue();
-    outV = NULL;
+    if (outV) outV->resize(0, 0);
     return;
   }
   MatrixPtr outV = getOutputValue();

paddle/gserver/layers/ExpandConvBaseLayer.cpp

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

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);
   }

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_;

paddle/gserver/layers/MKLDNNConvLayer.cpp

+/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "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

paddle/gserver/layers/MKLDNNConvLayer.h

+/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include "MKLDNNLayer.h"
+#include "mkldnn.hpp"
+
+namespace paddle {
+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

paddle/gserver/layers/MKLDNNFcLayer.cpp

@@ -17,9 +17,6 @@ limitations under the License. */
 
 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 {
 
@@ -93,35 +90,88 @@ void MKLDNNFcLayer::reshape(
   printSizeInfo();
 }
 
-void MKLDNNFcLayer::resetFwd(std::vector<mkldnn::primitive>& pipeline,
+void MKLDNNFcLayer::resetFwd(std::vector<primitive>& pipeline,
                              MKLDNNMatrixPtr& in,
                              MKLDNNMatrixPtr& wgt,
                              MKLDNNMatrixPtr& bias,
                              MKLDNNMatrixPtr& out) {
-  pipeline.clear();
-  bool hasBias = biases_ && biases_->getW();
-  const MatrixPtr& wgtVal = weight_->getW();
-  const MatrixPtr& biasVal = hasBias ? biases_->getW() : nullptr;
-  const MatrixPtr& outVal = output_.value;
+  resetFwdBuffers(in, wgt, bias, out);
+
+  resetFwdPD(fwdPD_, in, wgt, bias, out);
+
+  resetFwdPipeline(pipeline, fwdPD_, in, wgt, bias, out);
+
+  printValueFormatFlow();
+}
+
+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& inVal = getInputValue(0);
-    in = std::dynamic_pointer_cast<MKLDNNMatrix>(inVal);
+    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& inVal = getInputValue(0, CPU_DEVICE);
+    const MatrixPtr& cpuIn = getInputValue(0, CPU_DEVICE);
     in = MKLDNNMatrix::create(
-        inVal, memory::dims{bs_, ic_, ih_, iw_}, format::nchw, engine_);
+        cpuIn, {bs_, ic_, ih_, iw_}, format::nchw, engine_);
   }
   in->downSpatial();
+}
+
+void MKLDNNFcLayer::resetWgtBiasValue(MKLDNNMatrixPtr& wgt,
+                                      MKLDNNMatrixPtr& bias) {
   wgt = MKLDNNMatrix::create(
-      wgtVal, memory::dims{oc_, ic_, ih_, iw_}, format::oihw, engine_);
+      weight_->getW(), {oc_, ic_, ih_, iw_}, format::oihw, engine_);
   wgt->downSpatial();
-  bias = hasBias ? MKLDNNMatrix::create(biasVal, {oc_}, format::x, engine_)
-                 : nullptr;
-  out = MKLDNNMatrix::create(outVal, {bs_, oc_}, format::nc, engine_);
 
+  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>(out);
   if (!outputIsOnlyMKLDNN()) {
@@ -129,46 +179,59 @@ void MKLDNNFcLayer::resetFwd(std::vector<mkldnn::primitive>& pipeline,
     // 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,
-                                                in->getMemoryDesc(),
-                                                wgt->getMemoryDesc(),
-                                                bias->getMemoryDesc(),
-                                                out->getMemoryDesc())
-                                 : fc_fwd::desc(pk,
-                                                in->getMemoryDesc(),
-                                                wgt->getMemoryDesc(),
-                                                out->getMemoryDesc());
-  fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_);
-  if (hasBias) {
-    fwd_.reset(new fc_fwd(fwdPD, *in, *wgt, *bias, *out));
+  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, *in, *wgt, *out));
+    fwd_.reset(new fc_fwd(*pd, *in, *wgt, *out));
   }
-  printValueFormatFlow();
 
   pipeline.push_back(*fwd_);
 }
 
-void MKLDNNFcLayer::resetBwd(std::vector<mkldnn::primitive>& pipeline,
-                             MKLDNNMatrixPtr& in,
-                             MKLDNNMatrixPtr& wgt,
-                             MKLDNNMatrixPtr& bias,
-                             MKLDNNMatrixPtr& out) {
-  pipeline.clear();
-  if (!needResetBwd_) {
-    return;
-  }
-  needResetBwd_ = false;
-  bool hasBias = biases_ && biases_->getWGrad();
+void MKLDNNFcLayer::resetBwdBuffers(MKLDNNMatrixPtr& in,
+                                    MKLDNNMatrixPtr& wgt,
+                                    MKLDNNMatrixPtr& bias,
+                                    MKLDNNMatrixPtr& out) {
+  resetOutGrad(out);
+
+  resetWgtBiasGrad(wgt, bias);
 
-  /// backward weight
-  CHECK(inVal_) << "Should have input value";
-  const MatrixPtr& wgtGrad = weight_->getWGrad();
-  const MatrixPtr& biasGrad = hasBias ? biases_->getWGrad() : nullptr;
+  resetInGrad(in);
+}
 
+void MKLDNNFcLayer::resetOutGrad(MKLDNNMatrixPtr& out) {
   // TODO(TJ): merge outgrad
   int device = outputIsOnlyMKLDNN() ? MKLDNN_DEVICE : CPU_DEVICE;
   // for MKLDNN device:
@@ -178,66 +241,88 @@ void MKLDNNFcLayer::resetBwd(std::vector<mkldnn::primitive>& pipeline,
   // 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& outGrad = getOutput(device).grad;
-  out = MKLDNNMatrix::create(outGrad, outVal_->getPrimitiveDesc());
-  wgt = MKLDNNMatrix::create(wgtGrad, wgtVal_->getPrimitiveDesc());
-  bias = hasBias ? MKLDNNMatrix::create(biasGrad, biasVal_->getPrimitiveDesc())
-                 : nullptr;
-
-  // create memory primitive desc
-  fc_fwd::desc fwdDesc = fc_fwd::desc(prop_kind::forward,
-                                      inVal_->getMemoryDesc(),
-                                      wgt->getMemoryDesc(),
-                                      out->getMemoryDesc());
-  fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_);
-  fc_bwdWgt::desc bwdWgtDesc = hasBias
-                                   ? fc_bwdWgt::desc(inVal_->getMemoryDesc(),
-                                                     wgt->getMemoryDesc(),
-                                                     bias->getMemoryDesc(),
-                                                     out->getMemoryDesc())
-                                   : fc_bwdWgt::desc(inVal_->getMemoryDesc(),
-                                                     wgt->getMemoryDesc(),
-                                                     out->getMemoryDesc());
-  fc_bwdWgt::primitive_desc bwdWgtPD =
-      fc_bwdWgt::primitive_desc(bwdWgtDesc, engine_, fwdPD);
-
-  if (hasBias) {
-    bwdWgt_.reset(new fc_bwdWgt(bwdWgtPD, *inVal_, *out, *wgt, *bias));
-  } else {
-    bwdWgt_.reset(new fc_bwdWgt(bwdWgtPD, *inVal_, *out, *wgt));
+  CHECK(outVal_);
+  out =
+      MKLDNNMatrix::create(getOutput(device).grad, outVal_->getPrimitiveDesc());
+}
+
+void MKLDNNFcLayer::resetWgtBiasGrad(MKLDNNMatrixPtr& wgt,
+                                     MKLDNNMatrixPtr& bias) {
+  CHECK(wgtVal_);
+  wgt = MKLDNNMatrix::create(weight_->getWGrad(), wgtVal_->getPrimitiveDesc());
+
+  bias = nullptr;
+  if (biasVal_ == nullptr) {
+    return;
   }
-  pipeline.push_back(*bwdWgt_);
+  bias =
+      MKLDNNMatrix::create(biases_->getWGrad(), biasVal_->getPrimitiveDesc());
+}
 
-  /// backward data
+void MKLDNNFcLayer::resetInGrad(MKLDNNMatrixPtr& in) {
+  in = nullptr;
   const MatrixPtr& inGrad = inputLayers_[0]->getOutput().grad;
   if (inGrad == nullptr) {
     return;
   }
-  if (getInput(0, MKLDNN_DEVICE).getAllCount() > 1) {
-    // TODO(TJ): use outputMaps_ ways to get the inGrad_ when merge outgrad done
-  } else {
-    in = MKLDNNMatrix::create(inGrad, inVal_->getPrimitiveDesc());
-  }
-
-  fc_bwdData::desc bwdDataDesc = fc_bwdData::desc(
-      inVal_->getMemoryDesc(), wgt->getMemoryDesc(), out->getMemoryDesc());
-  fc_bwdData::primitive_desc bwdDataPD =
-      fc_bwdData::primitive_desc(bwdDataDesc, engine_, fwdPD);
+  // TODO(TJ): use outputMaps_ ways to get the inGrad_ when merge outgrad done
+  CHECK(inVal_);
+  in = MKLDNNMatrix::create(inGrad, inVal_->getPrimitiveDesc());
+}
 
-  CHECK(wgtVal_) << "Should have weight memory";
-  bwdData_.reset(new fc_bwdData(bwdDataPD, *out, *wgtVal_, *in));
-  printGradFormatFlow();
-  pipeline.push_back(*bwdData_);
+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::updateInputData() {
-  inVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
+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_));
 }
 
-void MKLDNNFcLayer::updateWeights(const UpdateCallback& callback) {
-  weight_->getParameterPtr()->incUpdate(callback);
-  if (biases_ && biases_->getWGrad()) {
-    biases_->getParameterPtr()->incUpdate(callback);
+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_);
+
+  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

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_;
@@ -67,6 +73,59 @@ public:
   void convertWeightsFromPaddle() override;
 
   void convertWeightsToPaddle() override;
+
+protected:
+  /**
+   * Forward functions: reset buffers(input, output, weight and bias),
+   *                    reset primitive descriptor,
+   *                    reset pipeline.
+   */
+  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);
+
+  /**
+   * Backward functions: reset buffers(input, output, weight and bias),
+   *                     reset primitive descriptor for backward weight,
+   *                     reset primitive descriptor for backward data,
+   *                     reset pipeline.
+   */
+  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

paddle/gserver/layers/MKLDNNPoolLayer.cpp

+/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "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

paddle/gserver/layers/MKLDNNPoolLayer.h

+/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include "MKLDNNLayer.h"
+#include "mkldnn.hpp"
+
+namespace paddle {
+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

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) {

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) {

paddle/math/MKLDNNMatrix.cpp

@@ -49,6 +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) {

paddle/math/MKLDNNMatrix.h

@@ -52,6 +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.

paddle/gserver/layers/ExpandConvBaseLayer.h → paddle/math/NEONFunctions.cpp

@@ -12,46 +12,44 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#pragma once
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
 
-#include <vector>
-#include "ConvBaseLayer.h"
-#include "paddle/math/Matrix.h"
+#include "NEONFunctions.h"
+#include <arm_neon.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 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

paddle/operators/concat_op.cu → paddle/math/NEONFunctions.h

@@ -4,7 +4,7 @@ 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
+    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,
@@ -12,8 +12,12 @@ 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/concat_op.h"
+#pragma once
 
-namespace ops = paddle::operators;
-// TODO(Yancey1989) Add GPU kernel
+namespace paddle {
+namespace neon {
+
+void relu(const float* a, float* b, int len);
+
+}  // namespace neon
+}  // namespace paddle

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);

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

paddle/operators/CMakeLists.txt

 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})

paddle/operators/accuracy_op.cc

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

paddle/operators/accuracy_op.cu

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

paddle/operators/accuracy_op.h

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

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());
   }
 };
 

paddle/operators/concat_op.cc

@@ -25,8 +25,11 @@ class ConcatOp : public framework::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::Tensor>("Out");
+    auto *out = ctx.Output<framework::LoDTensor>("Out");
     size_t axis = static_cast<size_t>(ctx.Attr<int>("axis"));
     size_t n = ins.size();
 

paddle/operators/cond_op.h

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

paddle/operators/elementwise_mul_op.cu

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

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());
   }
 };

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());