Merge branch 'develop' of https://github.com/baidu/Paddle into conv_op

cmake/external/mkldnn.cmake

@@ -51,7 +51,7 @@ ExternalProject_Add(
     ${EXTERNAL_PROJECT_LOG_ARGS}
     DEPENDS             ${MKLDNN_DEPENDS}
     GIT_REPOSITORY      "https://github.com/01org/mkl-dnn.git"
-    GIT_TAG             "v0.9"
+    GIT_TAG             "v0.10"
     PREFIX              ${MKLDNN_SOURCES_DIR}
     UPDATE_COMMAND      ""
     CMAKE_ARGS          -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR}

cmake/external/mklml.cmake

@@ -28,7 +28,7 @@ INCLUDE(ExternalProject)
 
 SET(MKLML_PROJECT       "extern_mklml")
 SET(MKLML_VER           "mklml_lnx_2018.0.20170720")
-SET(MKLML_URL           "https://github.com/01org/mkl-dnn/releases/download/v0.9/${MKLML_VER}.tgz")
+SET(MKLML_URL           "https://github.com/01org/mkl-dnn/releases/download/v0.10/${MKLML_VER}.tgz")
 SET(MKLML_SOURCE_DIR    "${THIRD_PARTY_PATH}/mklml")
 SET(MKLML_DOWNLOAD_DIR  "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}")
 SET(MKLML_DST_DIR       "mklml")
@@ -54,7 +54,8 @@ ExternalProject_Add(
     ${EXTERNAL_PROJECT_LOG_ARGS}
     PREFIX                ${MKLML_SOURCE_DIR}
     DOWNLOAD_DIR          ${MKLML_DOWNLOAD_DIR}
-    DOWNLOAD_COMMAND      wget --no-check-certificate -qO- ${MKLML_URL} | tar xz -C ${MKLML_DOWNLOAD_DIR}
+    DOWNLOAD_COMMAND      wget --no-check-certificate ${MKLML_URL} -c -q -O ${MKLML_VER}.tgz 
+                          && tar zxf ${MKLML_VER}.tgz
     DOWNLOAD_NO_PROGRESS  1
     UPDATE_COMMAND        ""
     CMAKE_ARGS            -DCMAKE_INSTALL_PREFIX=${MKLML_INSTALL_ROOT}

cmake/external/openblas.cmake

@@ -25,7 +25,12 @@ IF(NOT ${CBLAS_FOUND})
         "${CBLAS_INSTALL_DIR}/lib/${CMAKE_STATIC_LIBRARY_PREFIX}openblas${CMAKE_STATIC_LIBRARY_SUFFIX}"
         CACHE FILEPATH "openblas library." FORCE)
 
-    SET(COMMON_ARGS CC=${CMAKE_C_COMPILER} NO_SHARED=1 NO_LAPACK=1 libs)
+    IF(APPLE)
+        SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -isysroot ${CMAKE_OSX_SYSROOT}")
+        SET(COMMON_ARGS CC=${OPENBLAS_CC} NO_SHARED=1 NO_LAPACK=1 libs)
+    ELSE()
+        SET(COMMON_ARGS CC=${CMAKE_C_COMPILER} NO_SHARED=1 NO_LAPACK=1 libs)
+    ENDIF()
 
     IF(CMAKE_CROSSCOMPILING)
         IF(ANDROID)
@@ -40,11 +45,11 @@ IF(NOT ${CBLAS_FOUND})
             SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER} TARGET=${TARGET} ARM_SOFTFP_ABI=1 USE_THREAD=0)
         ELSEIF(RPI)
             # use hardfp
-            SET(OPENBLAS_COMMIT "v0.2.19")
+            SET(OPENBLAS_COMMIT "v0.2.20")
             SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER} TARGET=ARMV7 USE_THREAD=0)
         ENDIF()
     ELSE()
-        SET(OPENBLAS_COMMIT "v0.2.19")
+        SET(OPENBLAS_COMMIT "v0.2.20")
         SET(OPTIONAL_ARGS "")
         IF(CMAKE_SYSTEM_PROCESSOR MATCHES "^x86(_64)?$")
             SET(OPTIONAL_ARGS DYNAMIC_ARCH=1 NUM_THREADS=64)

doc/design/functions_operators_layers.md

+# Design Doc: Functions, Operators, and Layers
+
+In a DL system, we can compose one or more fine grained operators into a coarse grained one.  For example, the FC layer can be composed of a multiplication operator and an add operator.
+
+Historically, some fine grained operations are known as operators, and some coarse level ones are known as layers.  But we need a well-defined separation.
+
+In general, operators are those very fine grained operations, e.g., mul and add. In the implementation, we can write them as C++ functions:
+
+```c++
+template <typename T> T add(T x, T y) { return x + y; }
+template <typename T> T mul(T x, T y) { return x * y; }
+```
+
+Then we can wrap them into operators which are C++ classes and can be created from Python bindings by name.  A C macro can do this. For example, the following macro invocation
+
+```c++
+#define MAKE_FUNCTION_OPERATOR(mul);
+```
+
+generates
+
+```c++
+template <typename T> class mulOp : public OperatorBase {...};
+REGISTER_OP(mulOp<float32>, "mul");
+```
+
+so that in Python we can create operator mul by:
+
+```python
+X1 = Var()
+X2 = Var()
+Y = Var()
+paddle.cpp.create_operator("mul", input=[X1, X2], output=Y)
+```
+
+Also, at the same time, we can compose a coarse level C++ operator class by composing functions `mul` and `add`:
+
+```c++
+template <typename T>
+class FCOp : public OperatorBase {
+ public:
+  void Run(...) {
+    add(mul(Input<T>("X"), Input<T>("W")), Input<T>("b");
+  }
+};
+REGISTER_OP(FCOp, "fc");
+```
+
+We need to support such composition in Python as well.  To do so, we need a higher level Python wrapping of operator creation than `paddle.cpp.create_operator`.  This higher level operator API should be compatible with the layer API.
+
+Let's explain using an example.  Suppose that we are going to compose the FC using mul and add in Python, we'd like to have Python functions `mul` and `add` defined in module `operator`:
+
+```python
+def operator.mul(X1, X2):
+    O = Var()
+    paddle.cpp.create_operator("mul", input={X1, Y1], output=O)
+    return O
+
+def operator.add(X1, X2):
+    O = Var()
+    paddle.cpp.create_operator("add", input={X1, X2], output=O)
+    return O
+```
+
+Above code snippets are automatically generated.  Given them, users can define
+
+```python
+def layer.fc(X):
+    W = Var()
+    b = Var()
+    return operator.add(operator.mul(X, W), b)
+```
+
+If we don't have `operator.mul` and `operator.add`, the definiton of `layer.fc` would be complicated:
+
+```python
+def layer.fc(X):
+    W = Var()
+    b = Var()
+    O1 = Var()
+    paddle.cpp.create_operator("mul", input=[X, W], output=O1)
+    O2 = Var()
+    paddle.cpp.create_operator("add", input=[O1, b], output=O2)
+    return O2
+```
+
+We'd like to have Python bindings to operators in package `paddle.operator`, and Python compositions of operators in package `paddle.layer`.  So we have the following concepts in above illustrative example:
+
+```
+| C++ functions/functors | mul          | add          |             |          |
+| C++ operator class     | mulOp        | addOp        | FCOp        |          |
+| Python binding         | operator.mul | operator.add | operator.fc |          |
+| Python function        |              |              |             | layer.fc |
+```
+
+This is how we differentiate layer and operators in PaddlePaddle:
+
+- those defined in C++ and have a lightweighted Python wrapper in module `operators` are operators; whereas
+- those who don't have C++ implementations but a Python implementation that compose C++ operators are known as layers.

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:
+
+```python
+import paddle as pd
+
+x = var()
+y = var()
+cond = var()
+default_value = var()
+b = pd.create_ifelseop(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))
+
+with b.false_block():
+    x = b.inputs(0)
+    z = layer.fc(x)
+    b.set_output(0, operator.softmax(z))
+
+out = b(cond)
+```
+
+If only true_block is set in an IfElseOp, we can have a default value for false as:
+```python
+import paddle as pd
+
+x = var()
+y = var()
+cond = var()
+default_value = var()
+b = pd.create_ifelseop(inputs=[x], output_num=1, default_value)
+
+with b.true_block():
+    x = b.inputs(0)
+    z = operator.add(x, y)
+    b.set_output(0, operator.softmax(z))
+
+out = b(cond)
+```
+where default_value is a list of vars for `cond` == False.

doc/getstarted/build_and_install/index_cn.rst

@@ -6,14 +6,12 @@
 安装流程
 ++++++++
 
-PaddlePaddle提供数个预编译的二进制来进行安装，包括Docker镜像，ubuntu的deb安装包等。我们推荐使用Docker镜像来部署环境，同时欢迎贡献更多的安装包。
+PaddlePaddle提供Docker镜像来部署环境。
 
 .. toctree::
    :maxdepth: 1
    
    docker_install_cn.rst 
-   ubuntu_install_cn.rst
-
 
 
 编译流程

doc/getstarted/build_and_install/index_en.rst

@@ -8,14 +8,13 @@ Install PaddlePaddle
     :maxdepth: 1
 
     docker_install_en.rst
-    ubuntu_install_en.rst
 
 Build from Source
 -----------------
 
 ..  warning::
 
-    Please use :code:`deb` package or :code:`docker` image to install paddle. The building guide is used for hacking or contributing PaddlePaddle source code.
+    Please use :code:`docker` image to install paddle. The building guide is used for hacking or contributing PaddlePaddle source code.
 
 ..  toctree::
     :maxdepth: 1

doc/getstarted/build_and_install/ubuntu_install_cn.rst

-Ubuntu部署PaddlePaddle
-===================================
-
-PaddlePaddle提供了ubuntu 14.04 deb安装包。
-
-安装
-------
-
-安装包的下载地址是\: https://github.com/PaddlePaddle/Paddle/releases
-
-它包含四个版本\:
-
-* cpu版本: 支持主流x86处理器平台, 使用了avx指令集。
-
-* cpu-noavx版本：支持主流x86处理器平台，没有使用avx指令集。
-
-* gpu版本：支持主流x86处理器平台，支持nvidia cuda平台，使用了avx指令集。
-
-* gpu-noavx版本：支持主流x86处理器平台，支持nvidia cuda平台，没有使用avx指令集。
-
-下载完相关安装包后，执行:
-
-..  code-block:: shell
-
-    sudo apt-get install gdebi
-    gdebi paddle-*-cpu.deb
-
-或者:
-
-..  code-block:: shell
-
-    dpkg -i paddle-*-cpu.deb
-    apt-get install -f
-
-
-在 :code:`dpkg -i` 的时候如果报一些依赖未找到的错误是正常的，
-在 :code:`apt-get install -f` 里会继续安装 PaddlePaddle。
-
-安装完成后，可以使用命令 :code:`paddle version` 查看安装后的paddle 版本:
-
-..  code-block:: shell
-
-    PaddlePaddle 0.8.0b1, compiled with
-        with_avx: ON
-        with_gpu: OFF
-        with_double: OFF
-        with_python: ON
-        with_rdma: OFF
-        with_timer: OFF
-        with_predict_sdk:
-
-
-可能遇到的问题
---------------
-
-libcudart.so/libcudnn.so找不到
-++++++++++++++++++++++++++++++
-
-安装完成后，运行 :code:`paddle train` 报错\:
-
-..  code-block:: shell
-
-      0831 12:36:04.151525  1085 hl_dso_loader.cc:70] Check failed: nullptr != *dso_handle For Gpu version of PaddlePaddle, it couldn't find CUDA library: libcudart.so Please make sure you already specify its path.Note: for training data on Cpu using Gpu version of PaddlePaddle,you must specify libcudart.so via LD_LIBRARY_PATH.
-
-原因是未设置cuda运行时环境变量。 如果使用GPU版本的PaddlePaddle，请安装CUDA 7.5 和CUDNN 5到本地环境中，并设置：
-
-..  code-block:: shell
-
-    export LD_LIBRARY_PATH=/usr/local/cuda/lib64:/usr/local/cuda/lib:$LD_LIBRARY_PATH
-    export PATH=/usr/local/cuda/bin:$PATH
-

doc/getstarted/build_and_install/ubuntu_install_en.rst

-Debian Package installation guide
-=================================
-
-PaddlePaddle supports :code:`deb` pacakge. The installation of this :code:`deb` package is tested in ubuntu 14.04, but it should be support other debian based linux, too.
-
-There are four versions of debian package, :code:`cpu`, :code:`gpu`, :code:`cpu-noavx`, :code:`gpu-noavx`. And :code:`noavx` version is used to support CPU which does not contain :code:`AVX` instructions. The download url of :code:`deb` package is \: https://github.com/baidu/Paddle/releases/
-
-
-After downloading PaddlePaddle deb packages, you can use :code:`gdebi` install.
-
-..	code-block:: bash
-
-	gdebi paddle-*.deb
-
-If :code:`gdebi` is not installed, you can use :code:`sudo apt-get install gdebi` to install it.
-
-Or you can use following commands to install PaddlePaddle.
-
-..	code-block:: bash
-
-	dpkg -i paddle-*.deb
-	apt-get install -f
-
-And if you use GPU version deb package, you need to install CUDA toolkit and cuDNN, and set related environment variables(such as LD_LIBRARY_PATH) first. It is normal when `dpkg -i` get errors. `apt-get install -f` will continue install paddle, and install dependences. 
-

doc/howto/dev/new_op_cn.md

@@ -178,13 +178,13 @@ class MulKernel : public framework::OpKernel {
 
 ```c++
 namespace ops = paddle::operators;
-REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad);
+REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, ops::MulOpGrad);
 REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUPlace, float>);
 REGISTER_OP_CPU_KERNEL(mul_grad,
               ops::MulGradKernel<paddle::platform::CPUPlace, float>);
 ```
     
-  - `REGISTER_OP` ： 注册`ops::MulOp`类，类型名为`mul`，该类的`ProtoMaker`为`ops::MulOpMaker`，注册`ops::MulOpGrad`，类型名为`mul_grad`，
+  - `REGISTER_OP` ： 注册`ops::MulOp`类，类型名为`mul`，该类的`ProtoMaker`为`ops::MulOpMaker`，并且注册`ops::MulOpGrad`为其反向Op。
   - `REGISTER_OP_WITHOUT_GRADIENT` ： 用于注册没有反向的Op。
   - `REGISTER_OP_CPU_KERNEL` ：注册`ops::MulKernel`类，并特化模板参数为`paddle::platform::CPUPlace`和`float`类型，同理，注册`ops::MulKernel`类。
 

paddle/cuda/include/hl_cnn.h

@@ -173,6 +173,96 @@ extern void hl_avgpool_backward(const int frameCnt,
                                 real* backGrad,
                                 const int outStride);
 
+extern void hl_maxpool3D_forward(const int frameCnt,
+                                 const real* inputData,
+                                 const int channels,
+                                 const int depth,
+                                 const int height,
+                                 const int width,
+                                 const int pooledD,
+                                 const int pooledH,
+                                 const int pooledW,
+                                 const int sizeZ,
+                                 const int sizeY,
+                                 const int sizeX,
+                                 const int strideD,
+                                 const int strideH,
+                                 const int strideW,
+                                 const int paddingD,
+                                 const int paddingH,
+                                 const int paddingW,
+                                 real* tgtData,
+                                 real* maxPoolIdxData,
+                                 const int tgtStride);
+
+extern void hl_maxpool3D_backward(const int frameCnt,
+                                  const real* outGrad,
+                                  const int channels,
+                                  const int depth,
+                                  const int height,
+                                  const int width,
+                                  const int pooledD,
+                                  const int pooledH,
+                                  const int pooledW,
+                                  const int sizeZ,
+                                  const int sizeY,
+                                  const int sizeX,
+                                  const int strideD,
+                                  const int strideH,
+                                  const int strideW,
+                                  const int paddingD,
+                                  const int paddingH,
+                                  const int paddingW,
+                                  real scaleA,
+                                  real scaleB,
+                                  real* targetGrad,
+                                  real* maxPoolIdxData,
+                                  const int outStride);
+
+extern void hl_avgpool3D_forward(const int frameCnt,
+                                 const real* inputData,
+                                 const int channels,
+                                 const int depth,
+                                 const int height,
+                                 const int width,
+                                 const int pooledD,
+                                 const int pooledH,
+                                 const int pooledW,
+                                 const int sizeZ,
+                                 const int sizeY,
+                                 const int sizeX,
+                                 const int strideD,
+                                 const int strideH,
+                                 const int strideW,
+                                 const int paddingD,
+                                 const int paddingH,
+                                 const int paddingW,
+                                 real* tgtData,
+                                 const int tgtStride);
+
+extern void hl_avgpool3D_backward(const int frameCnt,
+                                  const real* outGrad,
+                                  const int channels,
+                                  const int depth,
+                                  const int height,
+                                  const int width,
+                                  const int pooledD,
+                                  const int pooledH,
+                                  const int pooledW,
+                                  const int sizeZ,
+                                  const int sizeY,
+                                  const int sizeX,
+                                  const int strideD,
+                                  const int strideH,
+                                  const int strideW,
+                                  int paddingD,
+                                  int paddingH,
+                                  int paddingW,
+                                  real scaleA,
+                                  real scaleB,
+                                  real* backGrad,
+                                  const int outStride);
+
 /**
  * @brief   Bilinear interpolation forward.
  *
@@ -275,4 +365,4 @@ extern void hl_maxout_backward(real* inGrad,
                                size_t featLen,
                                size_t groups);
 
-#endif /* HL_CNN_H_ */
+#endif  // HL_CNN_H_

paddle/cuda/include/hl_matrix.h

@@ -224,4 +224,80 @@ extern void hl_matrix_collect_shared_bias(real* B_d,
 extern void hl_matrix_rotate(
     real* mat, real* matRot, int dimM, int dimN, bool clockWise);
 
+/**
+ * @brief  Matrix vol2Col: Convert 3D volume into col matrix
+ *
+ * @param[in]   matSrc     input matrix.
+ * @param[in]   channel    channel of matSrc.
+ * @param[in]   depth      depth of matSrc.
+ * @param[in]   height     height of matSrc.
+ * @param[in]   width      width of matSrc.
+ * @param[in]   filterD    depth of filter.
+ * @param[in]   filterH    height of filter.
+ * @param[in]   filterW    width of filter.
+ * @param[in]   strideD    stride in the depth.
+ * @param[in]   strideH    stride in the height.
+ * @param[in]   strideW    stride in the width.
+ * @param[in]   paddingD   padding in the depth.
+ * @param[in]   paddingH   padding in the height.
+ * @param[in]   paddingW   padding in the width.
+ * @param[out]   dataDst     output matrix.
+ *
+ */
+extern void hl_matrix_vol2Col(const real* dataSrc,
+                              int channels,
+                              int depth,
+                              int height,
+                              int width,
+                              int filterD,
+                              int filterH,
+                              int filterW,
+                              int strideD,
+                              int strideH,
+                              int strideW,
+                              int paddingD,
+                              int paddingH,
+                              int paddingW,
+                              real* dataDst);
+
+/**
+ * @brief  Matrix col2Vol: Convert col matrix into 3D volume
+ *
+ * @param[out]  matDst     output matrix.
+ * @param[in]   channel    channel of matDst.
+ * @param[in]   depth      depth of matDst.
+ * @param[in]   height     height of matDst.
+ * @param[in]   width      width of matDst.
+ * @param[in]   filterD    depth of filter.
+ * @param[in]   filterH    height of filter.
+ * @param[in]   filterW    width of filter.
+ * @param[in]   strideD    stride in the depth.
+ * @param[in]   strideH    stride in the height.
+ * @param[in]   strideW    stride in the width.
+ * @param[in]   paddingD   padding in the depth.
+ * @param[in]   paddingH   padding in the height.
+ * @param[in]   paddingW   padding in the width.
+ * @param[in]   matSrc     input matrix.
+ * @param[in]   beta       input
+ * @param[in]   alpha      input
+ *
+ */
+extern void hl_matrix_col2Vol(real* dataDst,
+                              int channels,
+                              int depth,
+                              int height,
+                              int width,
+                              int filterD,
+                              int filterH,
+                              int filterW,
+                              int strideD,
+                              int strideH,
+                              int strideW,
+                              int paddingD,
+                              int paddingH,
+                              int paddingW,
+                              const real* dataSrc,
+                              real alpha,
+                              real beta);
+
 #endif /* HL_MATRIX_H_ */

paddle/cuda/include/stub/hl_cnn_stub.h

@@ -87,6 +87,96 @@ inline void hl_avgpool_backward(const int frameCnt,
                                 real* backGrad,
                                 const int outStride) {}
 
+inline void hl_maxpool3D_forward(const int frameCnt,
+                                 const real* inputData,
+                                 const int channels,
+                                 const int depth,
+                                 const int height,
+                                 const int width,
+                                 const int pooledD,
+                                 const int pooledH,
+                                 const int pooledW,
+                                 const int sizeZ,
+                                 const int sizeY,
+                                 const int sizeX,
+                                 const int strideD,
+                                 const int strideH,
+                                 const int strideW,
+                                 const int paddingD,
+                                 const int paddingH,
+                                 const int paddingW,
+                                 real* tgtData,
+                                 real* maxPoolIdxData,
+                                 const int tgtStride) {}
+
+inline void hl_maxpool3D_backward(const int frameCnt,
+                                  const real* outGrad,
+                                  const int channels,
+                                  const int depth,
+                                  const int height,
+                                  const int width,
+                                  const int pooledD,
+                                  const int pooledH,
+                                  const int pooledW,
+                                  const int sizeZ,
+                                  const int sizeY,
+                                  const int sizeX,
+                                  const int strideD,
+                                  const int strideH,
+                                  const int strideW,
+                                  const int paddingD,
+                                  const int paddingH,
+                                  const int paddingW,
+                                  real scaleA,
+                                  real scaleB,
+                                  real* targetGrad,
+                                  real* maxPoolIdxData,
+                                  const int outStride) {}
+
+inline void hl_avgpool3D_forward(const int frameCnt,
+                                 const real* inputData,
+                                 const int channels,
+                                 const int depth,
+                                 const int height,
+                                 const int width,
+                                 const int pooledD,
+                                 const int pooledH,
+                                 const int pooledW,
+                                 const int sizeZ,
+                                 const int sizeY,
+                                 const int sizeX,
+                                 const int strideD,
+                                 const int strideH,
+                                 const int strideW,
+                                 const int paddingD,
+                                 const int paddingH,
+                                 const int paddingW,
+                                 real* tgtData,
+                                 const int tgtStride) {}
+
+inline void hl_avgpool3D_backward(const int frameCnt,
+                                  const real* outGrad,
+                                  const int channels,
+                                  const int depth,
+                                  const int height,
+                                  const int width,
+                                  const int pooledD,
+                                  const int pooledH,
+                                  const int pooledW,
+                                  const int sizeZ,
+                                  const int sizeY,
+                                  const int sizeX,
+                                  const int strideD,
+                                  const int strideH,
+                                  const int strideW,
+                                  const int paddingD,
+                                  const int paddingH,
+                                  const int paddingW,
+                                  real scaleA,
+                                  real scaleB,
+                                  real* backGrad,
+                                  const int outStride) {}
+
 inline void hl_bilinear_forward(const real* inData,
                                 const size_t inImgH,
                                 const size_t inImgW,

paddle/cuda/include/stub/hl_matrix_stub.h

@@ -99,4 +99,38 @@ inline void hl_matrix_collect_shared_bias(real* B_d,
 inline void hl_matrix_rotate(
     real* mat, real* matRot, int dimM, int dimN, bool clockWise) {}
 
+inline void hl_matrix_vol2Col(const real* dataSrc,
+                              int channels,
+                              int depth,
+                              int height,
+                              int width,
+                              int filterD,
+                              int filterH,
+                              int filterW,
+                              int strideD,
+                              int strideH,
+                              int strideW,
+                              int paddingD,
+                              int paddingH,
+                              int paddingW,
+                              real* dataDst) {}
+
+inline void hl_matrix_col2Vol(real* dataDst,
+                              int channels,
+                              int depth,
+                              int height,
+                              int width,
+                              int filterD,
+                              int filterH,
+                              int filterW,
+                              int strideD,
+                              int strideH,
+                              int strideW,
+                              int paddingD,
+                              int paddingH,
+                              int paddingW,
+                              const real* dataSrc,
+                              real alpha,
+                              real beta) {}
+
 #endif  // HL_MATRIX_STUB_H_

paddle/cuda/src/hl_cuda_cnn.cu

@@ -353,6 +353,433 @@ void hl_avgpool_backward(const int frameCnt,
   CHECK_SYNC("hl_avgpool_backward failed");
 }
 
+__global__ void KeMaxPool3DForward(const int nthreads,
+                                   const real* inputData,
+                                   const int channels,
+                                   const int depth,
+                                   const int height,
+                                   const int width,
+                                   const int pooledD,
+                                   const int pooledH,
+                                   const int pooledW,
+                                   const int ksizeD,
+                                   const int ksizeH,
+                                   const int ksizeW,
+                                   const int strideD,
+                                   const int strideH,
+                                   const int strideW,
+                                   const int padD,
+                                   const int padH,
+                                   const int padW,
+                                   real* tgtData,
+                                   real* maxPoolIdxData,
+                                   const int tgtStride) {
+  for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < (nthreads);
+       index += blockDim.x * gridDim.x) {
+    int pw = index % pooledW;
+    int ph = (index / pooledW) % pooledH;
+    int pd = (index / pooledW / pooledH) % pooledD;
+    int c = (index / pooledW / pooledH / pooledD) % channels;
+    int frameNum = index / pooledW / pooledH / pooledD / channels;
+    int dstart = pd * strideD - padD;
+    int hstart = ph * strideH - padH;
+    int wstart = pw * strideW - padW;
+    int dend = min(dstart + ksizeD, depth);
+    int hend = min(hstart + ksizeH, height);
+    int wend = min(wstart + ksizeW, width);
+    dstart = max(dstart, 0);
+    hstart = max(hstart, 0);
+    wstart = max(wstart, 0);
+    real maxval = -FLT_MAX;
+    int maxIdx = -1;
+    inputData += (frameNum * channels + c) * depth * height * width;
+    for (int d = dstart; d < dend; ++d) {
+      for (int h = hstart; h < hend; ++h) {
+        for (int w = wstart; w < wend; ++w) {
+          if (maxval < inputData[(d * height + h) * width + w]) {
+            maxval = inputData[(d * height + h) * width + w];
+            maxIdx = (d * height + h) * width + w;
+          }
+        }
+      }
+    }
+    int tgtIndex =
+        index % (pooledW * pooledH * pooledD * channels) + frameNum * tgtStride;
+    tgtData[tgtIndex] = maxval;
+    maxPoolIdxData[tgtIndex] = maxIdx;
+  }
+}
+
+void hl_maxpool3D_forward(const int frameCnt,
+                          const real* inputData,
+                          const int channels,
+                          const int depth,
+                          const int height,
+                          const int width,
+                          const int pooledD,
+                          const int pooledH,
+                          const int pooledW,
+                          const int sizeZ,
+                          const int sizeY,
+                          const int sizeX,
+                          const int strideD,
+                          const int strideH,
+                          const int strideW,
+                          const int padD,
+                          const int padH,
+                          const int padW,
+                          real* tgtData,
+                          real* maxPoolIdxData,
+                          const int tgtStride) {
+  int num_kernels = pooledD * pooledH * pooledW * channels * frameCnt;
+  int blocks = (num_kernels + 1024 - 1) / 1024;
+  dim3 threads(1024, 1);
+  dim3 grid(blocks, 1);
+
+  KeMaxPool3DForward<<<grid, threads, 0, STREAM_DEFAULT>>>(num_kernels,
+                                                           inputData,
+                                                           channels,
+                                                           depth,
+                                                           height,
+                                                           width,
+                                                           pooledD,
+                                                           pooledH,
+                                                           pooledW,
+                                                           sizeZ,
+                                                           sizeY,
+                                                           sizeX,
+                                                           strideD,
+                                                           strideH,
+                                                           strideW,
+                                                           padD,
+                                                           padH,
+                                                           padW,
+                                                           tgtData,
+                                                           maxPoolIdxData,
+                                                           tgtStride);
+  CHECK_SYNC("hl_maxpool3D_forward failed");
+}
+
+__global__ void KeMaxPool3DBackward(const int nthreads,
+                                    const real* outGrad,
+                                    const int channels,
+                                    const int depth,
+                                    const int height,
+                                    const int width,
+                                    const int pooledD,
+                                    const int pooledH,
+                                    const int pooledW,
+                                    const int sizeZ,
+                                    const int sizeY,
+                                    const int sizeX,
+                                    const int strideD,
+                                    const int strideH,
+                                    const int strideW,
+                                    const int padD,
+                                    const int padH,
+                                    const int padW,
+                                    real scaleA,
+                                    real scaleB,
+                                    real* targetGrad,
+                                    real* maxPoolIdxData,
+                                    const int outStride) {
+  for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < (nthreads);
+       index += blockDim.x * gridDim.x) {
+    int offsetW = index % width;
+    int offsetH = (index / width) % height;
+    int offsetD = (index / width / height) % depth;
+    int offsetC = (index / width / height / depth) % channels;
+    int frameNum = index / width / height / depth / channels;
+
+    int pdstart =
+        (offsetD + padD < sizeZ) ? 0 : (offsetD + padD - sizeZ) / strideD + 1;
+    int phstart =
+        (offsetH + padH < sizeY) ? 0 : (offsetH + padH - sizeY) / strideH + 1;
+    int pwstart =
+        (offsetW + padW < sizeX) ? 0 : (offsetW + padW - sizeX) / strideW + 1;
+    int pdend = min((offsetD + padD) / strideD + 1, pooledD);
+    int phend = min((offsetH + padH) / strideH + 1, pooledH);
+    int pwend = min((offsetW + padW) / strideW + 1, pooledW);
+
+    real gradient = 0;
+    outGrad += ((frameNum * channels + offsetC) * pooledD * pooledH * pooledW);
+    maxPoolIdxData +=
+        ((frameNum * channels + offsetC) * pooledD * pooledH * pooledW);
+    for (int pd = pdstart; pd < pdend; ++pd) {
+      for (int ph = phstart; ph < phend; ++ph) {
+        for (int pw = pwstart; pw < pwend; ++pw) {
+          if (((offsetD * height + offsetH) * width + offsetW) ==
+              maxPoolIdxData[(pd * pooledH + ph) * pooledW + pw])
+            gradient += outGrad[(pd * pooledH + ph) * pooledW + pw];
+        }
+      }
+    }
+    targetGrad[index] = scaleA * gradient + scaleB * targetGrad[index];
+  }
+}
+
+void hl_maxpool3D_backward(const int frameCnt,
+                           const real* outGrad,
+                           const int channels,
+                           const int depth,
+                           const int height,
+                           const int width,
+                           const int outputD,
+                           const int outputH,
+                           const int outputW,
+                           const int sizeZ,
+                           const int sizeY,
+                           const int sizeX,
+                           const int strideD,
+                           const int strideH,
+                           const int strideW,
+                           const int paddingD,
+                           const int paddingH,
+                           const int paddingW,
+                           real scaleA,
+                           real scaleB,
+                           real* targetGrad,
+                           real* maxPoolIdxData,
+                           const int outStride) {
+  int num_kernels = depth * height * width * channels * frameCnt;
+  int blocks = (num_kernels + 1024 - 1) / 1024;
+
+  KeMaxPool3DBackward<<<blocks, 1024, 0, STREAM_DEFAULT>>>(num_kernels,
+                                                           outGrad,
+                                                           channels,
+                                                           depth,
+                                                           height,
+                                                           width,
+                                                           outputD,
+                                                           outputH,
+                                                           outputW,
+                                                           sizeZ,
+                                                           sizeY,
+                                                           sizeX,
+                                                           strideD,
+                                                           strideH,
+                                                           strideW,
+                                                           paddingD,
+                                                           paddingH,
+                                                           paddingW,
+                                                           scaleA,
+                                                           scaleB,
+                                                           targetGrad,
+                                                           maxPoolIdxData,
+                                                           outStride);
+  CHECK_SYNC("hl_maxpool3D_backward");
+}
+
+__global__ void KeAvgPool3DForward(const int nthreads,
+                                   const real* inputData,
+                                   const int channels,
+                                   const int depth,
+                                   const int height,
+                                   const int width,
+                                   const int pooledD,
+                                   const int pooledH,
+                                   const int pooledW,
+                                   const int sizeZ,
+                                   const int sizeY,
+                                   const int sizeX,
+                                   const int strideD,
+                                   const int strideH,
+                                   const int strideW,
+                                   const int padD,
+                                   const int padH,
+                                   const int padW,
+                                   real* tgtData,
+                                   const int tgtStride) {
+  for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < (nthreads);
+       index += blockDim.x * gridDim.x) {
+    int pw = index % pooledW;
+    int ph = (index / pooledW) % pooledH;
+    int pd = (index / pooledW / pooledH) % pooledD;
+    int c = (index / pooledW / pooledH / pooledD) % channels;
+    int frameNum = index / pooledW / pooledH / pooledD / channels;
+    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);
+    dstart = max(dstart, 0);
+    hstart = max(hstart, 0);
+    wstart = max(wstart, 0);
+    dend = min(dend, depth);
+    hend = min(hend, height);
+    wend = min(wend, width);
+
+    real aveval = 0;
+    inputData += (frameNum * channels + c) * depth * height * width;
+    for (int d = dstart; d < dend; ++d) {
+      for (int h = hstart; h < hend; ++h) {
+        for (int w = wstart; w < wend; ++w) {
+          aveval += inputData[(d * height + h) * width + w];
+        }
+      }
+    }
+    int tgtIndex =
+        index % (pooledW * pooledH * pooledD * channels) + frameNum * tgtStride;
+    tgtData[tgtIndex] = aveval / pool_size;
+  }
+}
+
+void hl_avgpool3D_forward(const int frameCnt,
+                          const real* inputData,
+                          const int channels,
+                          const int depth,
+                          const int height,
+                          const int width,
+                          const int pooledD,
+                          const int pooledH,
+                          const int pooledW,
+                          const int sizeZ,
+                          const int sizeY,
+                          const int sizeX,
+                          const int strideD,
+                          const int strideH,
+                          const int strideW,
+                          const int paddingD,
+                          const int paddingH,
+                          const int paddingW,
+                          real* tgtData,
+                          const int tgtStride) {
+  int num_kernels = pooledD * pooledH * pooledW * channels * frameCnt;
+  int blocks = (num_kernels + 1024 - 1) / 1024;
+  KeAvgPool3DForward<<<blocks, 1024, 0, STREAM_DEFAULT>>>(num_kernels,
+                                                          inputData,
+                                                          channels,
+                                                          depth,
+                                                          height,
+                                                          width,
+                                                          pooledD,
+                                                          pooledH,
+                                                          pooledW,
+                                                          sizeZ,
+                                                          sizeY,
+                                                          sizeX,
+                                                          strideD,
+                                                          strideH,
+                                                          strideW,
+                                                          paddingD,
+                                                          paddingH,
+                                                          paddingW,
+                                                          tgtData,
+                                                          tgtStride);
+  CHECK_SYNC("hl_avgpool3D_forward failed");
+}
+
+__global__ void KeAvgPool3DBackward(const int nthreads,
+                                    const real* outGrad,
+                                    const int channels,
+                                    const int depth,
+                                    const int height,
+                                    const int width,
+                                    const int pooledD,
+                                    const int pooledH,
+                                    const int pooledW,
+                                    const int sizeZ,
+                                    const int sizeY,
+                                    const int sizeX,
+                                    const int strideD,
+                                    const int strideH,
+                                    const int strideW,
+                                    const int padD,
+                                    const int padH,
+                                    const int padW,
+                                    real scaleA,
+                                    real scaleB,
+                                    real* tgtGrad,
+                                    const int outStride) {
+  for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < (nthreads);
+       index += blockDim.x * gridDim.x) {
+    int offsetW = index % width + padW;
+    int offsetH = (index / width) % height + padH;
+    int offsetD = (index / width / height) % depth + padD;
+    int offsetC = (index / width / height / depth) % channels;
+    int frameNum = index / width / height / depth / channels;
+
+    int pdstart = (offsetD < sizeZ) ? 0 : (offsetD - sizeZ) / strideD + 1;
+    int phstart = (offsetH < sizeY) ? 0 : (offsetH - sizeY) / strideH + 1;
+    int pwstart = (offsetW < sizeX) ? 0 : (offsetW - sizeX) / strideW + 1;
+    int pdend = min(offsetD / strideD + 1, pooledD);
+    int phend = min(offsetH / strideH + 1, pooledH);
+    int pwend = min(offsetW / strideW + 1, pooledW);
+
+    real gradient = 0;
+    outGrad += (frameNum * channels + offsetC) * pooledD * pooledH * pooledW;
+
+    for (int pd = pdstart; pd < pdend; ++pd) {
+      for (int ph = phstart; ph < phend; ++ph) {
+        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 poolsize = (dend - dstart) * (hend - hstart) * (wend - wstart);
+          gradient += outGrad[(pd * pooledH + ph) * pooledW + pw] / poolsize;
+        }
+      }
+    }
+    tgtGrad[index] = scaleA * gradient + scaleB * tgtGrad[index];
+  }
+}
+
+void hl_avgpool3D_backward(const int frameCnt,
+                           const real* outGrad,
+                           const int channels,
+                           const int depth,
+                           const int height,
+                           const int width,
+                           const int outputD,
+                           const int outputH,
+                           const int outputW,
+                           const int sizeZ,
+                           const int sizeY,
+                           const int sizeX,
+                           const int strideD,
+                           const int strideH,
+                           const int strideW,
+                           int paddingD,
+                           int paddingH,
+                           int paddingW,
+                           real scaleA,
+                           real scaleB,
+                           real* backGrad,
+                           const int outStride) {
+  int num_kernels = depth * height * width * channels * frameCnt;
+  int blocks = (num_kernels + 1024 - 1) / 1024;
+
+  KeAvgPool3DBackward<<<blocks, 1024, 0, STREAM_DEFAULT>>>(num_kernels,
+                                                           outGrad,
+                                                           channels,
+                                                           depth,
+                                                           height,
+                                                           width,
+                                                           outputD,
+                                                           outputH,
+                                                           outputW,
+                                                           sizeZ,
+                                                           sizeY,
+                                                           sizeX,
+                                                           strideD,
+                                                           strideH,
+                                                           strideW,
+                                                           paddingD,
+                                                           paddingH,
+                                                           paddingW,
+                                                           scaleA,
+                                                           scaleB,
+                                                           backGrad,
+                                                           outStride);
+  CHECK_SYNC("hl_avgpool3D_backward failed");
+}
+
 __global__ void KeBilinearInterpFw(const real* in,
                                    const size_t inImgH,
                                    const size_t inImgW,

paddle/cuda/src/hl_cuda_matrix.cu

@@ -592,3 +592,204 @@ void hl_matrix_rotate(
       mat, matRot, dimM, dimN, clockWise);
   CHECK_SYNC("hl_matrix_rotate failed");
 }
+
+__global__ void keMatrixVol2Col(int num_kernels,
+                                const real* dataSrc,
+                                real* dataDst,
+                                int depth,
+                                int height,
+                                int width,
+                                int filterD,
+                                int filterH,
+                                int filterW,
+                                int strideD,
+                                int strideH,
+                                int strideW,
+                                int paddingD,
+                                int paddingH,
+                                int paddingW,
+                                int depth_col,
+                                int height_col,
+                                int width_col) {
+  for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < num_kernels;
+       index += blockDim.x * gridDim.x) {
+    int w_out = index % width_col;
+    int h_out = (index / width_col) % height_col;
+    int d_out = (index / width_col / height_col) % depth_col;
+    int channel_in = index / width_col / height_col / depth_col;
+    int channel_out = channel_in * filterD * filterH * filterW;
+    int w_in = w_out * strideW - paddingW;
+    int h_in = h_out * strideH - paddingH;
+    int d_in = d_out * strideD - paddingD;
+
+    dataDst +=
+        ((channel_out * depth_col + d_out) * height_col + h_out) * width_col +
+        w_out;
+    dataSrc += ((channel_in * depth + d_in) * height + h_in) * width + w_in;
+    for (int k = 0; k < filterD; ++k) {
+      for (int i = 0; i < filterH; ++i) {
+        for (int j = 0; j < filterW; ++j) {
+          int d = d_in + k;
+          int h = h_in + i;
+          int w = w_in + j;
+          *dataDst = (d >= 0 && d < depth && h >= 0 && h < height && w >= 0 &&
+                      w < width)
+                         ? dataSrc[(k * height + i) * width + j]
+                         : 0;
+          dataDst += depth_col * height_col * width_col;
+        }
+      }
+    }
+  }
+}
+
+void hl_matrix_vol2Col(const real* dataSrc,
+                       int channels,
+                       int depth,
+                       int height,
+                       int width,
+                       int filterD,
+                       int filterH,
+                       int filterW,
+                       int strideD,
+                       int strideH,
+                       int strideW,
+                       int paddingD,
+                       int paddingH,
+                       int paddingW,
+                       real* dataDst) {
+  int depth_col = (depth + 2 * paddingD - filterD) / strideD + 1;
+  int height_col = (height + 2 * paddingH - filterH) / strideH + 1;
+  int width_col = (width + 2 * paddingW - filterW) / strideW + 1;
+  int num_kernels = channels * depth_col * height_col * width_col;
+
+  const int threads = 512;
+  const int blocks = DIVUP(num_kernels, threads);
+
+  keMatrixVol2Col<<<blocks, threads, 0, STREAM_DEFAULT>>>(num_kernels,
+                                                          dataSrc,
+                                                          dataDst,
+                                                          depth,
+                                                          height,
+                                                          width,
+                                                          filterD,
+                                                          filterH,
+                                                          filterW,
+                                                          strideD,
+                                                          strideH,
+                                                          strideW,
+                                                          paddingD,
+                                                          paddingH,
+                                                          paddingW,
+                                                          depth_col,
+                                                          height_col,
+                                                          width_col);
+  CHECK_SYNC("hl_matrix_vol2Col failed");
+}
+
+__global__ void keMatrixCol2Vol(int num_kernels,
+                                real* dataDst,
+                                const real* dataSrc,
+                                int depth,
+                                int height,
+                                int width,
+                                int filterD,
+                                int filterH,
+                                int filterW,
+                                int strideD,
+                                int strideH,
+                                int strideW,
+                                int paddingD,
+                                int paddingH,
+                                int paddingW,
+                                int depth_col,
+                                int height_col,
+                                int width_col,
+                                real alpha,
+                                real beta) {
+  for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < num_kernels;
+       index += blockDim.x * gridDim.x) {
+    real srcVal = 0;
+    real dstVal = dataDst[index];
+    int w = index % width + paddingW;
+    int h = (index / width) % height + paddingH;
+    int d = (index / width / height) % depth + paddingD;
+    int c = index / width / height / depth;
+    // compute the start and end of the output
+    int w_col_start = (w < filterW) ? 0 : (w - filterW) / strideW + 1;
+    int w_col_end = min(w / strideW + 1, width_col);
+    int h_col_start = (h < filterH) ? 0 : (h - filterH) / strideH + 1;
+    int h_col_end = min(h / strideH + 1, height_col);
+    int d_col_start = (d < filterD) ? 0 : (d - filterD) / strideD + 1;
+    int d_col_end = min(d / strideD + 1, depth_col);
+
+    int offset = (c * filterD * filterW * filterH + d * filterW * filterH +
+                  h * filterW + w) *
+                 depth_col * height_col * width_col;
+
+    int coeff_d_col =
+        (1 - strideD * filterW * filterH * depth_col) * height_col * width_col;
+    int coeff_h_col =
+        (1 - strideH * filterW * depth_col * height_col) * width_col;
+    int coeff_w_col = (1 - strideW * depth_col * height_col * width_col);
+
+    for (int d_col = d_col_start; d_col < d_col_end; ++d_col) {
+      for (int h_col = h_col_start; h_col < h_col_end; ++h_col) {
+        for (int w_col = w_col_start; w_col < w_col_end; ++w_col) {
+          srcVal += dataSrc[offset + d_col * coeff_d_col + h_col * coeff_h_col +
+                            w_col * coeff_w_col];
+        }
+      }
+    }
+    dataDst[index] = alpha * srcVal + beta * dstVal;
+  }
+}
+
+void hl_matrix_col2Vol(real* dataDst,
+                       int channels,
+                       int depth,
+                       int height,
+                       int width,
+                       int filterD,
+                       int filterH,
+                       int filterW,
+                       int strideD,
+                       int strideH,
+                       int strideW,
+                       int paddingD,
+                       int paddingH,
+                       int paddingW,
+                       const real* dataSrc,
+                       real alpha,
+                       real beta) {
+  int depth_col = (depth + 2 * paddingD - filterD) / strideD + 1;
+  int height_col = (height + 2 * paddingH - filterH) / strideH + 1;
+  int width_col = (width + 2 * paddingW - filterW) / strideW + 1;
+  int num_kernels = channels * depth * height * width;
+
+  const int threads = 512;
+  const int blocks = DIVUP(num_kernels, threads);
+
+  keMatrixCol2Vol<<<blocks, threads, 0, STREAM_DEFAULT>>>(num_kernels,
+                                                          dataDst,
+                                                          dataSrc,
+                                                          depth,
+                                                          height,
+                                                          width,
+                                                          filterD,
+                                                          filterH,
+                                                          filterW,
+                                                          strideD,
+                                                          strideH,
+                                                          strideW,
+                                                          paddingD,
+                                                          paddingH,
+                                                          paddingW,
+                                                          depth_col,
+                                                          height_col,
+                                                          width_col,
+                                                          alpha,
+                                                          beta);
+
+  CHECK_SYNC("hl_matrix_col2Vol failed");
+}

paddle/framework/backward.md

@@ -18,7 +18,7 @@ A backward network is built up with several backward operators. Backward operato
 For example, we have got a `mul_op`, and we can register it's information and corresponding backward operator by the following macro:
 
 ```cpp
-REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad);
+REGISTER_OP(mul, MulOp, MulOpMaker, MulOpGrad);
 ```
 
 `mul` is the operator's type. `MulOp` and `MulOpMaker` are the operator class and the operator maker class respectively.

paddle/framework/backward_test.cc

@@ -127,8 +127,8 @@ class FillZeroOpMaker : public OpProtoAndCheckerMaker {
  public:
   FillZeroOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("x", "x");
-    AddOutput("out", "out");
+    AddInput("Src", "x");
+    AddOutput("Dst", "out");
     AddComment("");
   }
 };
@@ -138,7 +138,7 @@ class AddOpMaker : public OpProtoAndCheckerMaker {
   AddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "x").AsDuplicable();
-    AddOutput("Y", "y");
+    AddOutput("Out", "out");
     AddComment("");
   }
 };
@@ -148,16 +148,14 @@ class AddOpMaker : public OpProtoAndCheckerMaker {
 namespace f = paddle::framework;
 namespace ops = paddle::operators;
 using EnforceNotMet = paddle::platform::EnforceNotMet;
-REGISTER_OP(rowwise_add, f::NOP, f::RowWiseAddOpMaker, rowwise_add_grad,
-            f::NOP);
-REGISTER_OP(mul, f::NOP, f::MulOpMaker, mul_grad, f::NOP);
-REGISTER_OP(sigmoid, f::NOP, f::SigmoidOpMaker, sigmoid_grad, f::NOP);
+REGISTER_OP(rowwise_add, f::NOP, f::RowWiseAddOpMaker, f::NOP);
+REGISTER_OP(mul, f::NOP, f::MulOpMaker, f::NOP);
+REGISTER_OP(sigmoid, f::NOP, f::SigmoidOpMaker, f::NOP);
 REGISTER_OP_WITHOUT_GRADIENT(nograd, f::NOP, f::NoGradOpMaker);
 REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, f::NOP, f::FillZeroOpMaker);
-REGISTER_OP(add, f::NOP, f::AddOpMaker, add_grad, f::NOP);
+REGISTER_OP(add, f::NOP, f::AddOpMaker, f::NOP);
 REGISTER_OP_WITHOUT_GRADIENT(fc, f::FcOp, f::FcOpMaker);
-REGISTER_OP(many_output_op, f::NOP, f::ManyOutputOpMaker, many_output_op_grad,
-            f::NOP);
+REGISTER_OP(many_output_op, f::NOP, f::ManyOutputOpMaker, f::NOP);
 
 TEST(Backward, simple_op_grad) {
   auto fwd = f::OpRegistry::CreateOp(

paddle/framework/grad_op_builder_test.cc

@@ -54,8 +54,8 @@ TEST(GradOpBuilder, AddTwo) {
   EXPECT_EQ(grad_add_op->Output(f::GradVarName("Y")), f::GradVarName("y"));
 }
 
-REGISTER_OP(mult_io, f::NOP, f::MutiInOutOpMaker, mult_io_grad, f::NOP);
-REGISTER_OP(io_ignored, f::NOP, f::IOIgnoredOpMaker, io_ignored_grad, f::NOP);
+REGISTER_OP(mult_io, f::NOP, f::MutiInOutOpMaker, f::NOP);
+REGISTER_OP(io_ignored, f::NOP, f::IOIgnoredOpMaker, f::NOP);
 
 TEST(GradOpBuilder, MutiInOut) {
   std::shared_ptr<f::OperatorBase> test_op(f::OpRegistry::CreateOp(

paddle/framework/lod_tensor.cc

@@ -19,25 +19,24 @@
 namespace paddle {
 namespace framework {
 
-LODTensor::LOD LODTensor::LOD::SliceLevels(size_t level_begin,
-                                           size_t level_end) const {
+LOD SliceLevels(const LOD& in, size_t level_begin, size_t level_end) {
   LOD new_lod;
   new_lod.reserve(level_end - level_begin);
   for (size_t i = level_begin; i < level_end; i++) {
-    new_lod.emplace_back(at(i));
+    new_lod.emplace_back(in.at(i));
   }
   return new_lod;
 }
 
-LODTensor::LOD LODTensor::LOD::SliceInLevel(size_t level, size_t elem_begin,
-                                            size_t elem_end) const {
+LOD SliceInLevel(const LOD& in, size_t level, size_t elem_begin,
+                 size_t elem_end) {
   // slice the lod.
   LOD new_lod;
-  new_lod.reserve(size() - level);
-  auto start = this->at(level)[elem_begin];
-  auto end = this->at(level)[elem_end];
+  new_lod.reserve(in.size() - level);
+  auto start = in.at(level)[elem_begin];
+  auto end = in.at(level)[elem_end];
 
-  for (auto it = this->begin() + level; it != this->end(); it++) {
+  for (auto it = in.begin() + level; it != in.end(); it++) {
     auto it_begin = std::find(it->begin(), it->end(), start);
     auto it_end = std::find(it_begin, it->end(), end);
     PADDLE_ENFORCE(it_begin != it->end(), "error in parsing lod info");
@@ -49,11 +48,11 @@ LODTensor::LOD LODTensor::LOD::SliceInLevel(size_t level, size_t elem_begin,
                    [start](int v) { return v - start; });
     PADDLE_ENFORCE_EQ(new_lod.back().front(), 0, "error in slice LOD");
   }
-  PADDLE_ENFORCE_LE(new_lod.size(), this->size());
+  PADDLE_ENFORCE_LE(new_lod.size(), in.size());
   return new_lod;
 }
 
-bool operator==(const LODTensor::LOD& a, const LODTensor::LOD& b) {
+bool operator==(const LOD& a, const LOD& b) {
   if (a.size() != b.size()) {
     return false;
   }
@@ -70,9 +69,27 @@ bool operator==(const LODTensor::LOD& a, const LODTensor::LOD& b) {
       }
     }
   }
-
   return true;
 }
 
+void LODTensor::SliceLevels(size_t level_begin, size_t level_end) {
+  auto new_lod = framework::SliceLevels(lod_, level_begin, level_end);
+  lod_ = new_lod;
+}
+
+void LODTensor::SliceInLevel(size_t level, size_t elem_begin, size_t elem_end) {
+  PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
+                 NumLevels());
+  PADDLE_ENFORCE(elem_begin < NumElements(level),
+                 "element begin [%d] out of range [%d]", elem_begin,
+                 NumElements(level));
+  PADDLE_ENFORCE(elem_end < NumElements(level) + 1,
+                 "element end [%d] out of range [%d]", elem_end,
+                 NumElements(level));
+
+  auto new_lod = framework::SliceInLevel(lod_, level, elem_begin, elem_end);
+  lod_ = new_lod;
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/lod_tensor.h

@@ -15,7 +15,7 @@
 #pragma once
 
 #include <memory>
-#if !defined(PADDLE_ONLY_CPU)
+#ifndef PADDLE_ONLY_CPU
 #include <thrust/device_vector.h>
 #include <thrust/host_vector.h>
 #endif
@@ -27,33 +27,39 @@
 namespace paddle {
 namespace framework {
 
+#ifdef PADDLE_ONLY_CPU
+template <typename T>
+using Vector = std::vector<T>;
+#else
+template <typename T>
+using Vector = thrust::host_vector<T>;
+#endif
+
+using LOD = std::vector<Vector<size_t>>;
+
+LOD SliceLevels(const LOD& in, size_t level_begin, size_t level_end);
+
+LOD SliceInLevel(const LOD& in, size_t level, size_t elem_begin,
+                 size_t elem_end);
+
+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 : public Tensor {
+class LODTensor {
  public:
-// Level save offsets of each unit.
-#ifdef PADDLE_ONLY_CPU
-  template <typename T>
-  using Vector = std::vector<T>;
-#else
-  template <typename T>
-  using Vector = thrust::host_vector<T>;
-#endif
-  // LoD stores offsets of each level of units, the largest units level first,
-  // then the smaller units level. Each Level stores the offsets of units in
-  // Tesor.
-  class LOD : public std::vector<Vector<size_t>> {
-   public:
-    LOD SliceLevels(size_t level_begin, size_t level_end) const;
-    LOD SliceInLevel(size_t level, size_t elem_begin, size_t elem_end) const;
-  };
-
   LODTensor() {}
-  explicit LODTensor(const LOD &lod) : lod_(lod) {}
+  LODTensor(const LOD& lod, Tensor* t) : lod_(lod), tensor_(t) {}
+
+  void set_lod(const LOD& lod) { lod_ = lod; }
 
-  virtual Tensor *Clone() const { return new LODTensor(lod_); }
+  void set_tensor(Tensor* tensor) { tensor_ = tensor; }
+
+  Tensor& tensor() { return *tensor_; }
+
+  LOD lod() { return lod_; }
 
   /*
    * Get a element from LOD.
@@ -79,71 +85,23 @@ class LODTensor : public Tensor {
     PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
                    NumLevels());
     // the last offset is the end of last element
-    return lod_[level].size() - 1;
+    return (lod_)[level].size() - 1;
   }
 
   /*
-   * Slice of levels[level_begin:level_end], with tensor shared.
+   * Slice of levels[level_begin:level_end]
    */
-  template <typename T>
-  LODTensor SliceLevels(size_t level_begin, size_t level_end) const;
+  void SliceLevels(size_t level_begin, size_t level_end);
 
   /*
-   * Slice of elements of a level, [elem_begin: elem_end], with tensor shared.
+   * Slice of elements of a level, [elem_begin: elem_end]
    * @note: low performance in slice lod_.
    */
-  template <typename T>
-  LODTensor SliceInLevel(size_t level, size_t elem_begin,
-                         size_t elem_end) const;
-
-  /*
-   * Copy other's lod_'s content, free to mutate.
-   */
-  void CopyLOD(const LODTensor &other) { lod_ = other.lod_; }
-  /*
-   * Determine whether LODTensor has a valid LOD info.
-   */
-  const LOD &lod() const { return lod_; }
-  LOD *mutable_lod() { return &lod_; }
-
-  virtual ~LODTensor() {}
+  void SliceInLevel(size_t level, size_t elem_begin, size_t elem_end);
 
  private:
   LOD lod_;
+  Tensor* tensor_;  // not owned
 };
-
-bool operator==(const LODTensor::LOD &a, const LODTensor::LOD &b);
-
-template <typename T>
-LODTensor LODTensor::SliceLevels(size_t level_begin, size_t level_end) const {
-  auto new_lod = lod_.SliceLevels(level_begin, level_end);
-  // slice levels just need to update LOD info, each level will contains the
-  // whole tensor_, so no need to modify tensor_.
-  LODTensor new_tensor(new_lod);
-  new_tensor.ShareDataWith<T>(*this);
-  return new_tensor;
-}
-
-template <typename T>
-LODTensor LODTensor::SliceInLevel(size_t level, size_t elem_begin,
-                                  size_t elem_end) const {
-  PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
-                 NumLevels());
-  PADDLE_ENFORCE(elem_begin < NumElements(level),
-                 "element begin [%d] out of range [%d]", elem_begin,
-                 NumElements(level));
-  PADDLE_ENFORCE(elem_end < NumElements(level) + 1,
-                 "element end [%d] out of range [%d]", elem_end,
-                 NumElements(level));
-
-  auto new_lod = lod_.SliceInLevel(level, elem_begin, elem_end);
-
-  // slice elements just need to update LOD info, because offsets are not
-  // changed, so the original tensor_ can be reused.
-  LODTensor new_tensor(new_lod);
-  new_tensor.ShareDataWith<T>(*this);
-  return new_tensor;
-}
-
 }  // namespace framework
 }  // namespace paddle

paddle/framework/lod_tensor.md

+# Design Doc: LoD (Level-of-Detail) Tensor
+
+PaddlePaddle's RNN doesn't require that all instances have the same length.  To do so, we introduce an extension to Tensor, namely, LoD Tensor.
+
+## Challenge of Variable-length Inputs
+
+People usually represent a mini-batch by a Tensor.  For example, a mini-batch of 32 images, each of size 32x32, is a 10x32x32 Tensor.  So a transformation, T, of all images can be a matrix multiplication of the 32x32xO-dimensional tensor T and the 10x32x32 Tensor.
+
+Another example is that each mini-batch contains 32 sentences, where each word is a D-dimensional one-hot vector.  If all sentences have the same length L, we can represent this mini-batch by a 32xLxD tensor.  However, in most cases, sentences have variable lengths, and we will need an index data structure to record these variable lengths.
+
+## LoD as a Solution
+
+### Mini-Batch of variable-length sentenses
+
+Let's imagine a mini-batch of 3 variable lengths sentences, containing 3, 1, and 2 words respectively.  We can represent it by a (3+1+2)xD tensor plus some index information:
+
+```
+   3
+3   1 2
+||| | ||
+```
+
+Each `|` represents a D-dimensional word vectors.  The number 3 on top indicate 3 sentences, and numbers 3, 1, and 2 on the second level represent the number of words in each sentence.
+
+### Mini-Batch of variable-length videos
+
+This approach generalizes to the case where elements are not words, but higher dimensional objects, like images.  Suppose that a mini-batch contains videos of the same frame size 640x480.  If a mini-batch contains 3 videos of 3, 1, and 2 frames respectively.  The underlying tensor is of size (3+1+2)x640x480.  The index information illustrates as:
+
+```
+     3
+3     1  2
+口口口 口 口口
+```
+
+where each `口` represents an image.
+
+### Mini-Batch of fixed-size images
+
+Let's get back to a typical example, image classification, where each mini-batch has M fixed-sized images.  The LoD Tensor representation is
+
+```
+     M
+1 1 1 1     1
+口口口口 ... 口
+```
+
+The many 1's on the second level seem duplicated.  For this particular case of 2 levels and the second level always have length 1, we can ignore the LoD index.
+
+### Design and summarization
+
+In summary, as long as that the essential elements (words  or images) have the same size, we can represent mini-batches by a LoD Tensor:
+
+- The underlying tensor has size LxD1xD2x..., where D1xD2... is the size of the essential elements, and
+- the first dimension size L has an additon property -- a LoD index as a nested vector:
+
+  ```c++
+  typedef std::vector<std::vector> > LoD;
+  ```
+
+- The LoD index can is not necessary when there are only two levels and all elements of the second level have length 1.
+
+## Slicing of LoD Tensor
+
+Consider that we have a network with three levels of RNN: the top level one handles articles, the second level one handles sentences, and the basic level one handles words.  This network requires that mini-batches represented by 4 level LoD Tensor, for example,
+
+```
+         3
+3           1  2
+3   2  4    1  2  3
+||| || |||| |  || |||
+```
+
+To allow each level of RNN to handle its input, we define **the slicing of a LoD Tensor is defined as getting the j-th sequence on level i, or the <i,j>-slice**
+
+For example, the <2,1>-slice of above slice is
+
+```
+2
+||
+```
+
+and the <1,2>-slice of above example is
+
+```
+2
+2  3
+|| |||
+```
+
+Let's go on slicing this slice.  Its <1,1>-slice is
+
+```
+3
+|||
+```
+
+### The General Slicing Algorithm
+
+The algorithm, with over-simplified data structure, is defined as
+
+```c++
+typedef vector<vector<int> > LoD;
+
+struct LoDTensor {
+  LoD lod_;
+  float* tensor_;
+};
+
+LoDTensor Slice(const LoDTensor& lodt, int level, int sequence) {
+
+}
+```
+
+### Slicing the Top Level
+
+Please be aware that an RNN operator only slices the top level of a LoD Tensor to get the step inputs.
+
+```c++
+LoDTensor Slice(const LoDTensor& lodt, int sequence) {
+
+}
+```

paddle/framework/lod_tensor_test.cc

@@ -24,13 +24,12 @@ namespace framework {
 class LODTensorTester : public ::testing::Test {
  public:
   virtual void SetUp() override {
-    lod_tensor.reset(new LODTensor);
     // tensor's batch_size: 30
     // 3 levels
     // 0 10 20
     // 0 5 10 15 20
     // 0 2 5 7 10 12 15 20
-    LODTensor::LOD lod;
+    LOD lod;
     lod.push_back(std::vector<size_t>{0, 10, 20});
     lod.push_back(std::vector<size_t>{0, 5, 10, 15, 20});
     lod.push_back(std::vector<size_t>{0, 2, 5, 7, 10, 12, 15, 17, 20});
@@ -41,75 +40,65 @@ class LODTensorTester : public ::testing::Test {
     // malloc memory
     tensor.mutable_data<float>(place);
 
-    lod_tensor.reset(new LODTensor(lod));
-    lod_tensor->Resize({20 /*batch size*/, 128 /*dim*/});
-
-    lod_tensor->ShareDataWith<float>(tensor);
-    // lod_tensor->ShareDataWith<Tensor>(tensor);
+    lod_tensor.set_lod(lod);
+    lod_tensor.set_tensor(&tensor);
   }
 
  protected:
-  std::unique_ptr<LODTensor> lod_tensor;
   platform::CPUPlace place;
   Tensor 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) {
-    auto new_lod_tensor = lod_tensor->SliceLevels<float>(level, level + 1);
+    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(0UL), lod_tensor->NumElements(level));
-    // ASSERT_EQ(new_lod_tensor, *lod_tensor);
+    ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor.NumElements(level));
+    ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
+              lod_tensor.tensor().data<float>());
   }
   // slice 2 level
   for (size_t level = 0; level < 2UL; ++level) {
-    auto new_lod_tensor = lod_tensor->SliceLevels<float>(level, level + 2);
+    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.data<float>(), lod_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.tensor().data<float>(),
+              lod_tensor.tensor().data<float>());
   }
 }
 
 TEST_F(LODTensorTester, SliceInLevel) {
   size_t level = 0;
-  auto new_lod_tensor = lod_tensor->SliceInLevel<float>(level, 0, 2);
+  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.data<float>(), lod_tensor->data<float>());
+  ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
+            lod_tensor.tensor().data<float>());
 
   level = 1;
-  new_lod_tensor = lod_tensor->SliceInLevel<float>(level, 0, 2);
+  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.data<float>(), lod_tensor->data<float>());
-}
-
-TEST_F(LODTensorTester, ShareLOD) {
-  LODTensor new_lod_tensor;
-  new_lod_tensor.CopyLOD(*lod_tensor);
-  ASSERT_EQ(new_lod_tensor.lod(), lod_tensor->lod());
-}
-
-TEST_F(LODTensorTester, CopyLOD) {
-  LODTensor new_lod_tensor;
-  new_lod_tensor.CopyLOD(*lod_tensor);
-  bool equals = std::equal(lod_tensor->lod().begin(), lod_tensor->lod().end(),
-                           new_lod_tensor.lod().begin());
-  ASSERT_TRUE(equals);
+  ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
+            lod_tensor.tensor().data<float>());
 }
 
 }  // namespace framework

paddle/framework/op_info.h

@@ -80,9 +80,19 @@ class OpInfoMap {
   }
 
   const OpInfo& Get(const std::string& type) const {
+    auto op_info_ptr = GetNullable(type);
+    PADDLE_ENFORCE_NOT_NULL(op_info_ptr, "Operator %s has not been registered",
+                            type);
+    return *op_info_ptr;
+  }
+
+  const OpInfo* GetNullable(const std::string& type) const {
     auto it = map_.find(type);
-    PADDLE_ENFORCE(it != map_.end(), "Operator %s are not found", type);
-    return it->second;
+    if (it == map_.end()) {
+      return nullptr;
+    } else {
+      return &it->second;
+    }
   }
 
   template <typename Callback>

paddle/framework/op_registry.h

@@ -33,8 +33,7 @@ namespace framework {
 class OpRegistry {
  public:
   template <typename OpType, typename ProtoMakerType, typename GradOpType>
-  static void RegisterOp(const std::string& op_type,
-                         const std::string& grad_op_type) {
+  static void RegisterOp(const std::string& op_type) {
     PADDLE_ENFORCE(!OpInfoMap::Instance().Has(op_type),
                    "'%s' is registered more than once.", op_type);
     OpInfo op_info;
@@ -43,9 +42,9 @@ class OpRegistry {
         const VariableNameMap& outputs, const AttributeMap& attrs) {
       return new OpType(type, inputs, outputs, attrs);
     };
-    op_info.grad_op_type_ = grad_op_type;
     if (std::type_index(typeid(ProtoMakerType)) !=
         std::type_index(typeid(NOPMaker))) {
+      op_info.grad_op_type_ = op_type + "_grad";
       op_info.proto_ = new OpProto;
       op_info.checker_ = new OpAttrChecker;
       auto maker = ProtoMakerType(op_info.proto_, op_info.checker_);
@@ -55,15 +54,14 @@ class OpRegistry {
           op_info.proto_->IsInitialized(),
           "Fail to initialize %s's OpProto, because %s is not initialized",
           op_type, op_info.proto_->InitializationErrorString());
+      // register gradient op
+      RegisterOp<GradOpType, NOPMaker, NOP>(op_info.grad_op_type_);
     } else {
+      op_info.grad_op_type_ = "";
       op_info.proto_ = nullptr;
       op_info.checker_ = nullptr;
     }
     OpInfoMap::Instance().Insert(op_type, op_info);
-    // register gradient op
-    if (!grad_op_type.empty()) {
-      RegisterOp<GradOpType, NOPMaker, NOP>(grad_op_type, "");
-    }
   }
 
   static std::unique_ptr<OperatorBase> CreateOp(const std::string& type,
@@ -92,10 +90,8 @@ class Registrar {
 template <typename OpType, typename ProtoMakerType, typename GradOpType>
 class OpRegistrar : public Registrar {
  public:
-  explicit OpRegistrar(const char* op_type) { OpRegistrar(op_type, ""); }
-  OpRegistrar(const char* op_type, const char* grad_op_type) {
-    OpRegistry::RegisterOp<OpType, ProtoMakerType, GradOpType>(op_type,
-                                                               grad_op_type);
+  explicit OpRegistrar(const char* op_type) {
+    OpRegistry::RegisterOp<OpType, ProtoMakerType, GradOpType>(op_type);
   }
 };
 
@@ -121,8 +117,7 @@ class OpKernelRegistrar : public Registrar {
 /**
  * Macro to register Operator.
  */
-#define REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type,          \
-                    grad_op_class)                                            \
+#define REGISTER_OP(op_type, op_class, op_maker_class, grad_op_class)         \
   STATIC_ASSERT_GLOBAL_NAMESPACE(                                             \
       __reg_op__##op_type, "REGISTER_OP must be called in global namespace"); \
   class _OpClass_##op_type##_ : public op_class {                             \
@@ -137,14 +132,14 @@ class OpKernelRegistrar : public Registrar {
   };                                                                          \
   static ::paddle::framework::OpRegistrar<                                    \
       _OpClass_##op_type##_, op_maker_class, _OpGradClass_##op_type##_>       \
-      __op_registrar_##op_type##__(#op_type, #grad_op_type);                  \
+      __op_registrar_##op_type##__(#op_type);                                 \
   int TouchOpRegistrar_##op_type() {                                          \
     __op_registrar_##op_type##__.Touch();                                     \
     return 0;                                                                 \
   }
 
 #define REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class) \
-  REGISTER_OP(op_type, op_class, op_maker_class, , ::paddle::framework::NOP)
+  REGISTER_OP(op_type, op_class, op_maker_class, ::paddle::framework::NOP)
 
 /**
  * Macro to register OperatorKernel.

paddle/framework/operator.cc

@@ -33,12 +33,12 @@ ExecutionContext::GetEigenDevice<platform::GPUPlace, Eigen::GpuDevice>() const {
 }
 #endif
 
-const std::string& OperatorBase::Input(const std::string& name) const {
+std::string OperatorBase::Input(const std::string& name) const {
   auto& ins = Inputs(name);
-  PADDLE_ENFORCE_EQ(ins.size(), 1UL,
+  PADDLE_ENFORCE_LE(ins.size(), 1UL,
                     "Op %s input %s should contain only one variable", type_,
                     name);
-  return ins[0];
+  return ins.empty() ? kEmptyVarName : ins[0];
 }
 
 const std::vector<std::string>& OperatorBase::Inputs(
@@ -49,12 +49,12 @@ const std::vector<std::string>& OperatorBase::Inputs(
   return it->second;
 }
 
-const std::string& OperatorBase::Output(const std::string& name) const {
+std::string OperatorBase::Output(const std::string& name) const {
   auto& outs = Outputs(name);
-  PADDLE_ENFORCE_EQ(outs.size(), 1UL,
+  PADDLE_ENFORCE_LE(outs.size(), 1UL,
                     "Op %s output %s should contain only one variable", type_,
                     name);
-  return outs[0];
+  return outs.empty() ? kEmptyVarName : outs[0];
 }
 
 const std::vector<std::string>& OperatorBase::Outputs(
@@ -119,16 +119,8 @@ OperatorBase::OperatorBase(const std::string& type,
                            const VariableNameMap& outputs,
                            const AttributeMap& attrs)
     : type_(type), inputs_(inputs), outputs_(outputs), attrs_(attrs) {
-  static std::atomic<size_t> gUniqId(0UL);
-  for (auto& output : outputs_) {
-    for (auto& output_name : output.second) {
-      if (output_name == kTempVarName) {
-        output_name += type_;
-        output_name += "@";
-        output_name += std::to_string(gUniqId.fetch_add(1));
-      }
-    }
-  }
+  GenerateTemporaryNames();
+  CheckAllInputOutputSet();
 }
 
 std::vector<std::string> OperatorBase::OutputVars(bool has_intermediate) const {
@@ -156,6 +148,35 @@ std::vector<std::string> OperatorBase::OutputVars(bool has_intermediate) const {
   return ret_val;
 }
 
+void OperatorBase::CheckAllInputOutputSet() const {
+  auto& info_map = OpInfoMap::Instance();
+  auto* op_info = info_map.GetNullable(Type());
+  if (op_info == nullptr || op_info->proto_ == nullptr) return;
+
+  for (auto& in : op_info->Proto().inputs()) {
+    PADDLE_ENFORCE(inputs_.find(in.name()) != inputs_.end(),
+                   "Type %s's input %s is not set", Type(), in.name());
+  }
+
+  for (auto& out : op_info->Proto().outputs()) {
+    PADDLE_ENFORCE(outputs_.find(out.name()) != outputs_.end(),
+                   "Type %s's output %s is not set", Type(), out.name());
+  }
+}
+
+void OperatorBase::GenerateTemporaryNames() {
+  static std::atomic<size_t> gUniqId(0UL);
+  for (auto& output : outputs_) {
+    for (auto& output_name : output.second) {
+      if (output_name == kTempVarName) {
+        output_name += type_;
+        output_name += "@";
+        output_name += std::to_string(gUniqId.fetch_add(1));
+      }
+    }
+  }
+}
+
 void OpProtoAndCheckerMaker::Validate() {
   validated_ = true;
   CheckNoDuplicatedInOutAttrs();

paddle/framework/operator.h

@@ -95,12 +95,12 @@ class OperatorBase {
   const VariableNameMap& Inputs() const { return inputs_; }
   const VariableNameMap& Outputs() const { return outputs_; }
   //! Get a input with argument's name described in `op_proto`
-  const std::string& Input(const std::string& name) const;
+  std::string Input(const std::string& name) const;
   //! Get a input which has multiple variables.
   const std::vector<std::string>& Inputs(const std::string& name) const;
 
   //! Get a output with argument's name described in `op_proto`
-  const std::string& Output(const std::string& name) const;
+  std::string Output(const std::string& name) const;
   //! Get an output which has multiple variables.
   //! TODO add a vector_view to prevent memory copy.
   const std::vector<std::string>& Outputs(const std::string& name) const;
@@ -127,6 +127,10 @@ class OperatorBase {
   // IG (Inputs Gradients)
   VariableNameMap outputs_;
   AttributeMap attrs_;
+
+ private:
+  void GenerateTemporaryNames();
+  void CheckAllInputOutputSet() const;
 };
 
 // Macro for define a clone method.
@@ -238,11 +242,13 @@ class InferShapeContext {
   }
 
   const Variable* InputVar(const std::string& name) const {
-    return scope_.FindVar(op_.Input(name));
+    auto ipt = op_.Input(name);
+    return ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt);
   }
 
   Variable* OutputVar(const std::string& name) const {
-    return scope_.FindVar(op_.Output(name));
+    auto opt = op_.Output(name);
+    return opt == kEmptyVarName ? nullptr : scope_.FindVar(opt);
   }
 
   const std::vector<const Variable*> MultiInputVar(
@@ -250,9 +256,11 @@ class InferShapeContext {
     auto names = op_.Inputs(name);
     std::vector<const Variable*> res;
     res.reserve(names.size());
-    std::transform(
-        names.begin(), names.end(), std::back_inserter(res),
-        [this](const std::string& name) { return scope_.FindVar(name); });
+    std::transform(names.begin(), names.end(), std::back_inserter(res),
+                   [this](const std::string& name) {
+                     return name == kEmptyVarName ? nullptr
+                                                  : scope_.FindVar(name);
+                   });
     return res;
   }
 
@@ -260,24 +268,24 @@ class InferShapeContext {
     auto names = op_.Outputs(name);
     std::vector<const Variable*> res;
     res.reserve(names.size());
-    std::transform(
-        names.begin(), names.end(), std::back_inserter(res),
-        [this](const std::string& name) { return scope_.FindVar(name); });
+    std::transform(names.begin(), names.end(), std::back_inserter(res),
+                   [this](const std::string& name) {
+                     return name == kEmptyVarName ? nullptr
+                                                  : scope_.FindVar(name);
+                   });
     return res;
   }
 
   template <typename T>
   const T* Input(const std::string& name) const {
     auto* var = InputVar(name);
-    PADDLE_ENFORCE_NOT_NULL(var, "Input(%s) should not be nullptr", name);
-    return &var->Get<T>();
+    return var == nullptr ? nullptr : &var->Get<T>();
   }
 
   template <typename T>
   T* Output(const std::string& name) const {
     auto var = OutputVar(name);
-    PADDLE_ENFORCE_NOT_NULL(var, "Output(%s) should not be nullptr", name);
-    return var->GetMutable<T>();
+    return var == nullptr ? nullptr : var->GetMutable<T>();
   }
 
   template <typename T>
@@ -288,10 +296,7 @@ class InferShapeContext {
     std::transform(names.begin(), names.end(), std::back_inserter(res),
                    [&](const std::string& sub_name) {
                      auto var = scope_.FindVar(sub_name);
-                     PADDLE_ENFORCE_NOT_NULL(
-                         var, "MultiInput(%s:%s) should not be nullptr", name,
-                         sub_name);
-                     return &var->Get<T>();
+                     return var == nullptr ? nullptr : &var->Get<T>();
                    });
     return res;
   }
@@ -304,10 +309,7 @@ class InferShapeContext {
     std::transform(names.begin(), names.end(), std::back_inserter(res),
                    [&](const std::string& sub_name) {
                      auto var = scope_.FindVar(sub_name);
-                     PADDLE_ENFORCE_NOT_NULL(
-                         var, "MultiOutput(%s:%s) should not be nullptr.", name,
-                         sub_name);
-                     return var->GetMutable<T>();
+                     return var == nullptr ? nullptr : var->GetMutable<T>();
                    });
     return res;
   }

paddle/framework/tensor_impl.h

@@ -117,6 +117,8 @@ inline void Tensor::CopyFrom(const Tensor& src,
     memory::Copy(boost::get<platform::GPUPlace>(dst_place), dst_ptr,
                  boost::get<platform::GPUPlace>(src_place), src_ptr, size, 0);
   }
+  PADDLE_ENFORCE(cudaStreamSynchronize(0),
+                 "cudaStreamSynchronize failed in Tensor CopyFrom");
 
 #endif
 }

paddle/function/CMakeLists.txt

@@ -21,6 +21,8 @@ if(USE_NNPACK)
   endif()
 endif()
 
+list(APPEND cpp_files neon/NeonDepthwiseConv.cpp)
+
 add_library(paddle_function STATIC ${cpp_files} ${cu_objs})
 add_dependencies(paddle_function ${external_project_dependencies})
 add_dependencies(paddle_function paddle_proto)
@@ -42,11 +44,11 @@ if(WITH_GPU)
     add_simple_unittest(RowConvOpTest)
     add_simple_unittest(BlockExpandOpTest)
     add_simple_unittest(CropOpTest)
-    add_simple_unittest(DepthwiseConvOpTest)
 endif()
 
 add_simple_unittest(Im2ColTest)
 add_simple_unittest(GemmConvOpTest)
+add_simple_unittest(DepthwiseConvOpTest)
 endif()
 
 add_style_check_target(paddle_function ${h_files})

paddle/function/DepthwiseConvOpTest.cpp

@@ -34,4 +34,13 @@ TEST(DepthwiseConv, BackwardFilter) {
 }
 #endif
 
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+
+TEST(DepthwiseConv, Forward) {
+  DepthwiseConvolution<DEVICE_TYPE_CPU, DEVICE_TYPE_CPU>(
+      "GemmConv-CPU", "NeonDepthwiseConv-CPU", forward);
+}
+
+#endif
+
 }  // namespace paddle

paddle/function/Im2Col.h

@@ -16,6 +16,7 @@ limitations under the License. */
 
 #include "TensorShape.h"
 #include "TensorType.h"
+#include "neon/neon_util.h"
 
 namespace paddle {
 
@@ -93,4 +94,95 @@ public:
                   int paddingWidth);
 };
 
+template <class T>
+struct Padding {
+  static void run(const T* src,
+                  T* dest,
+                  int channels,
+                  int inputHeight,
+                  int inputWidth,
+                  int paddingHeight,
+                  int paddingWidth) {
+    const int destWidth = inputWidth + 2 * paddingWidth;
+    for (int c = 0; c < channels; c++) {
+      if (paddingHeight > 0) {
+        memset(dest, 0, destWidth * paddingHeight * sizeof(T));
+        dest += destWidth * paddingHeight;
+      }
+
+      for (int i = 0; i < inputHeight; i++) {
+        // padding head
+        for (int j = 0; j < paddingWidth; j++) {
+          *dest++ = T(0);
+        }
+
+        memcpy(dest, src, inputWidth * sizeof(T));
+        dest += inputWidth;
+        src += inputWidth;
+
+        // padding tail
+        for (int j = 0; j < paddingWidth; j++) {
+          *dest++ = T(0);
+        }
+      }
+
+      if (paddingHeight > 0) {
+        memset(dest, 0, destWidth * paddingHeight * sizeof(T));
+        dest += destWidth * paddingHeight;
+      }
+    }
+  }
+};
+
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+template <>
+struct Padding<float> {
+  static void run(const float* src,
+                  float* dest,
+                  int channels,
+                  int inputHeight,
+                  int inputWidth,
+                  int paddingHeight,
+                  int paddingWidth) {
+    const int destWidth = inputWidth + 2 * paddingWidth;
+    for (int c = 0; c < channels; c++) {
+      if (paddingHeight > 0) {
+        memset(dest, 0, destWidth * paddingHeight * sizeof(float));
+        dest += destWidth * paddingHeight;
+      }
+
+      for (int i = 0; i < inputHeight; i++) {
+        // padding head
+        for (int j = 0; j < paddingWidth; j++) {
+          *dest++ = float(0);
+        }
+
+        int step = inputWidth >> 2;
+        int remain = inputWidth & 3;
+        for (int s = 0; s < step; s++) {
+          float32x4_t s0 = vld1q_f32(src);
+          vst1q_f32(dest, s0);
+          src += 4;
+          dest += 4;
+        }
+        for (int r = 0; r < remain; r++) {
+          *dest++ = *src++;
+        }
+
+        // padding tail
+        for (int j = 0; j < paddingWidth; j++) {
+          *dest++ = float(0);
+        }
+      }
+
+      if (paddingHeight > 0) {
+        memset(dest, 0, destWidth * paddingHeight * sizeof(float));
+        dest += destWidth * paddingHeight;
+      }
+    }
+  }
+};
+
+#endif
+
 }  // namespace paddle

paddle/function/neon/NeonDepthwiseConv.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 "neon_util.h"
+#include "paddle/function/ConvOp.h"
+#include "paddle/function/Im2Col.h"
+
+namespace paddle {
+
+namespace neon {
+
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+
+template <int filterSize, int stride>
+struct DepthwiseConvKernel {};
+
+inline float32_t conv3x3(float32x4_t r0,
+                         float32x4_t r1,
+                         float32x4_t r2,
+                         float32x4_t k0,
+                         float32x4_t k1,
+                         float32x4_t k2) {
+  float32x4_t tmp;
+  tmp = vmulq_f32(r0, k0);
+  tmp = vmlaq_f32(tmp, r1, k1);
+  tmp = vmlaq_f32(tmp, r2, k2);
+  return vaddvq_f32(tmp);
+}
+
+inline float32_t conv4x4(float32x4_t r0,
+                         float32x4_t r1,
+                         float32x4_t r2,
+                         float32x4_t r3,
+                         float32x4_t k0,
+                         float32x4_t k1,
+                         float32x4_t k2,
+                         float32x4_t k3) {
+  float32x4_t tmp;
+  tmp = vmulq_f32(r0, k0);
+  tmp = vmlaq_f32(tmp, r1, k1);
+  tmp = vmlaq_f32(tmp, r2, k2);
+  tmp = vmlaq_f32(tmp, r3, k3);
+  return vaddvq_f32(tmp);
+}
+
+/**
+ * Each step calculates four elements of the output.
+ * First step:
+ *   R0[0, 1, 2, 3...] * K[0][0]
+ *   R0[1, 2, 3, 4...] * K[0][1]
+ *   R0[2, 3, 4, 5...] * K[0][2]
+ *   R1[0, 1, 2, 3...] * K[1][0]
+ *   R1[1, 2, 3, 4...] * K[1][1]
+ *   R1[2, 3, 4, 5...] * K[1][2]
+ *   R2[0, 1, 2, 3...] * K[2][0]
+ *   R2[1, 2, 3, 4...] * K[2][1]
+ * + R2[2, 3, 4, 5...] * K[2][2]
+ * ------------------------------
+ *     Output[0, 1, 2, 3]
+ */
+template <>
+struct DepthwiseConvKernel<3, 1> {
+  static void run(const float* inputData,
+                  const float* filterData,
+                  int inputHeight,
+                  int inputWidth,
+                  int outputChannels,
+                  int outputHeight,
+                  int outputWidth,
+                  int filterMultiplier,
+                  float* outputData) {
+    const int steps = outputWidth >> 2;
+    const int remain = outputWidth & 3;
+    for (int c = 0; c < outputChannels; c++, filterData += 9) {
+      // Load the filters
+      float32x4_t k[3];
+      k[0] = vld1q_f32(filterData);
+      k[1] = vld1q_f32(filterData + 3);
+      k[2] = vld1q_f32(filterData + 6);
+      k[0] = vsetq_lane_f32(0.f, k[0], 3);
+      k[1] = vsetq_lane_f32(0.f, k[1], 3);
+      k[2] = vsetq_lane_f32(0.f, k[2], 3);
+
+      const float* r0 =
+          inputData + (c / filterMultiplier) * (inputHeight * inputWidth);
+      const float* r1 = r0 + inputWidth;
+      const float* r2 = r0 + inputWidth * 2;
+      float32x4_t input[3][3];
+      for (int h = 0; h < outputHeight; h++) {
+        for (int s = 0; s < steps; s++) {
+          // Load the inputs
+          float32x4_t tmp;
+          input[0][0] = vld1q_f32(r0);
+          tmp = vld1q_f32(r0 + 4);
+          input[0][1] = vextq_f32(input[0][0], tmp, 1);
+          input[0][2] = vextq_f32(input[0][0], tmp, 2);
+          input[1][0] = vld1q_f32(r1);
+          tmp = vld1q_f32(r1 + 4);
+          input[1][1] = vextq_f32(input[1][0], tmp, 1);
+          input[1][2] = vextq_f32(input[1][0], tmp, 2);
+          input[2][0] = vld1q_f32(r2);
+          tmp = vld1q_f32(r2 + 4);
+          input[2][1] = vextq_f32(input[2][0], tmp, 1);
+          input[2][2] = vextq_f32(input[2][0], tmp, 2);
+
+          float32x4_t tmp1 = vdupq_n_f32(0.f);
+          float32x4_t tmp2 = vdupq_n_f32(0.f);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[0][0], k[0], 0);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[0][1], k[0], 1);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[0][2], k[0], 2);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[1][0], k[1], 0);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[1][1], k[1], 1);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[1][2], k[1], 2);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[2][0], k[2], 0);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[2][1], k[2], 1);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[2][2], k[2], 2);
+          tmp1 = vaddq_f32(tmp1, tmp2);
+
+          vst1q_f32(outputData, tmp1);
+          r0 += 4;
+          r1 += 4;
+          r2 += 4;
+          outputData += 4;
+        }
+
+        for (int r = 0; r < remain; r++) {
+          float32x4_t i0 = vld1q_f32(r0);
+          float32x4_t i1 = vld1q_f32(r1);
+          float32x4_t i2 = vld1q_f32(r2);
+          *outputData = conv3x3(i0, i1, i2, k[0], k[1], k[2]);
+          r0++;
+          r1++;
+          r2++;
+          outputData++;
+        }
+
+        r0 += 2;
+        r1 += 2;
+        r2 += 2;
+      }
+    }
+  }
+};
+
+/**
+ * Each step calculates four elements of the output.
+ * First step:
+ *   R0[0, 2, 4, 6...] * K[0][0]
+ *   R0[1, 3, 5, 7...] * K[0][1]
+ *   R0[2, 4, 6, 8...] * K[0][2]
+ *   R1[0, 2, 4, 6...] * K[1][0]
+ *   R1[1, 3, 5, 7...] * K[1][1]
+ *   R1[2, 4, 6, 8...] * K[1][2]
+ *   R2[0, 2, 4, 6...] * K[2][0]
+ *   R2[1, 3, 5, 7...] * K[2][1]
+ *   R2[2, 4, 6, 8...] * K[2][2]
+ * ------------------------------
+ *     Output[0, 1, 2, 3]
+ */
+template <>
+struct DepthwiseConvKernel<3, 2> {
+  static void run(const float* inputData,
+                  const float* filterData,
+                  int inputHeight,
+                  int inputWidth,
+                  int outputChannels,
+                  int outputHeight,
+                  int outputWidth,
+                  int filterMultiplier,
+                  float* outputData) {
+    const int steps = outputWidth >> 2;
+    const int remain = outputWidth & 3;
+    for (int c = 0; c < outputChannels; c++, filterData += 9) {
+      // Load the filters
+      float32x4_t k[3];
+      k[0] = vld1q_f32(filterData);
+      k[1] = vld1q_f32(filterData + 3);
+      k[2] = vld1q_f32(filterData + 6);
+      k[0] = vsetq_lane_f32(0.f, k[0], 3);
+      k[1] = vsetq_lane_f32(0.f, k[1], 3);
+      k[2] = vsetq_lane_f32(0.f, k[2], 3);
+
+      const float* start =
+          inputData + (c / filterMultiplier) * (inputHeight * inputWidth);
+      float32x4_t input[3][3];
+      for (int h = 0; h < outputHeight; h++) {
+        const float* r0 = start + 2 * h * inputWidth;
+        const float* r1 = start + (2 * h + 1) * inputWidth;
+        const float* r2 = start + (2 * h + 2) * inputWidth;
+        for (int s = 0; s < steps; s++) {
+          // Load the inputs
+          float32x4_t data1;
+          float32x4x2_t data2;
+
+          data2 = vld2q_f32(r0);
+          input[0][0] = data2.val[0];
+          input[0][1] = data2.val[1];
+          data1 = vld1q_f32(r0 + 8);
+          input[0][2] = vextq_f32(data2.val[0], data1, 1);
+
+          data2 = vld2q_f32(r1);
+          input[1][0] = data2.val[0];
+          input[1][1] = data2.val[1];
+          data1 = vld1q_f32(r1 + 8);
+          input[1][2] = vextq_f32(data2.val[0], data1, 1);
+
+          data2 = vld2q_f32(r2);
+          input[2][0] = data2.val[0];
+          input[2][1] = data2.val[1];
+          data1 = vld1q_f32(r2 + 8);
+          input[2][2] = vextq_f32(data2.val[0], data1, 1);
+
+          float32x4_t tmp1 = vdupq_n_f32(0.f);
+          float32x4_t tmp2 = vdupq_n_f32(0.f);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[0][0], k[0], 0);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[0][1], k[0], 1);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[0][2], k[0], 2);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[1][0], k[1], 0);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[1][1], k[1], 1);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[1][2], k[1], 2);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[2][0], k[2], 0);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[2][1], k[2], 1);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[2][2], k[2], 2);
+          tmp1 = vaddq_f32(tmp1, tmp2);
+
+          vst1q_f32(outputData, tmp1);
+          r0 += 8;
+          r1 += 8;
+          r2 += 8;
+          outputData += 4;
+        }
+
+        for (int r = 0; r < remain; r++) {
+          float32x4_t i0 = vld1q_f32(r0);
+          float32x4_t i1 = vld1q_f32(r1);
+          float32x4_t i2 = vld1q_f32(r2);
+          *outputData = conv3x3(i0, i1, i2, k[0], k[1], k[2]);
+          r0 += 2;
+          r1 += 2;
+          r2 += 2;
+          outputData++;
+        }
+      }
+    }
+  }
+};
+
+/**
+ * Each step calculates four elements of the output.
+ */
+template <>
+struct DepthwiseConvKernel<4, 1> {
+  static void run(const float* inputData,
+                  const float* filterData,
+                  int inputHeight,
+                  int inputWidth,
+                  int outputChannels,
+                  int outputHeight,
+                  int outputWidth,
+                  int filterMultiplier,
+                  float* outputData) {
+    const int steps = outputWidth >> 2;
+    const int remain = outputWidth & 3;
+    for (int c = 0; c < outputChannels; c++, filterData += 16) {
+      // Load the filters
+      float32x4_t k[4];
+      k[0] = vld1q_f32(filterData);
+      k[1] = vld1q_f32(filterData + 4);
+      k[2] = vld1q_f32(filterData + 8);
+      k[3] = vld1q_f32(filterData + 12);
+
+      const float* r0 =
+          inputData + (c / filterMultiplier) * (inputHeight * inputWidth);
+      const float* r1 = r0 + inputWidth;
+      const float* r2 = r0 + inputWidth * 2;
+      const float* r3 = r0 + inputWidth * 3;
+      float32x4_t input[4][4];
+      for (int h = 0; h < outputHeight; h++) {
+        for (int s = 0; s < steps; s++) {
+          // Load the inputs
+          float32x4_t tmp;
+          input[0][0] = vld1q_f32(r0);
+          tmp = vld1q_f32(r0 + 4);
+          input[0][1] = vextq_f32(input[0][0], tmp, 1);
+          input[0][2] = vextq_f32(input[0][0], tmp, 2);
+          input[0][3] = vextq_f32(input[0][0], tmp, 3);
+
+          input[1][0] = vld1q_f32(r1);
+          tmp = vld1q_f32(r1 + 4);
+          input[1][1] = vextq_f32(input[1][0], tmp, 1);
+          input[1][2] = vextq_f32(input[1][0], tmp, 2);
+          input[1][3] = vextq_f32(input[1][0], tmp, 3);
+
+          input[2][0] = vld1q_f32(r2);
+          tmp = vld1q_f32(r2 + 4);
+          input[2][1] = vextq_f32(input[2][0], tmp, 1);
+          input[2][2] = vextq_f32(input[2][0], tmp, 2);
+          input[2][3] = vextq_f32(input[2][0], tmp, 3);
+
+          input[3][0] = vld1q_f32(r3);
+          tmp = vld1q_f32(r3 + 4);
+          input[3][1] = vextq_f32(input[3][0], tmp, 1);
+          input[3][2] = vextq_f32(input[3][0], tmp, 2);
+          input[3][3] = vextq_f32(input[3][0], tmp, 3);
+
+          float32x4_t tmp1 = vdupq_n_f32(0.f);
+          float32x4_t tmp2 = vdupq_n_f32(0.f);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[0][0], k[0], 0);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[0][1], k[0], 1);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[0][2], k[0], 2);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[0][3], k[0], 3);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[1][0], k[1], 0);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[1][1], k[1], 1);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[1][2], k[1], 2);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[1][3], k[1], 3);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[2][0], k[2], 0);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[2][1], k[2], 1);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[2][2], k[2], 2);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[2][3], k[2], 3);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[3][0], k[3], 0);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[3][1], k[3], 1);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[3][2], k[3], 2);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[3][3], k[3], 3);
+          tmp1 = vaddq_f32(tmp1, tmp2);
+
+          vst1q_f32(outputData, tmp1);
+          r0 += 4;
+          r1 += 4;
+          r2 += 4;
+          r3 += 4;
+          outputData += 4;
+        }
+
+        for (int r = 0; r < remain; r++) {
+          float32x4_t i0 = vld1q_f32(r0);
+          float32x4_t i1 = vld1q_f32(r1);
+          float32x4_t i2 = vld1q_f32(r2);
+          float32x4_t i3 = vld1q_f32(r3);
+          *outputData = conv4x4(i0, i1, i2, i3, k[0], k[1], k[2], k[3]);
+          r0++;
+          r1++;
+          r2++;
+          r3++;
+          outputData++;
+        }
+
+        r0 += 3;
+        r1 += 3;
+        r2 += 3;
+        r3 += 3;
+      }
+    }
+  }
+};
+
+/**
+ * Each step calculates four elements of the output.
+ */
+template <>
+struct DepthwiseConvKernel<4, 2> {
+  static void run(const float* inputData,
+                  const float* filterData,
+                  int inputHeight,
+                  int inputWidth,
+                  int outputChannels,
+                  int outputHeight,
+                  int outputWidth,
+                  int filterMultiplier,
+                  float* outputData) {
+    const int steps = outputWidth >> 2;
+    const int remain = outputWidth & 3;
+    for (int c = 0; c < outputChannels; c++, filterData += 16) {
+      // Load the filters
+      float32x4_t k[4];
+      k[0] = vld1q_f32(filterData);
+      k[1] = vld1q_f32(filterData + 4);
+      k[2] = vld1q_f32(filterData + 8);
+      k[3] = vld1q_f32(filterData + 12);
+
+      const float* start =
+          inputData + (c / filterMultiplier) * (inputHeight * inputWidth);
+      float32x4_t input[4][4];
+      for (int h = 0; h < outputHeight; h++) {
+        const float* r0 = start + 2 * h * inputWidth;
+        const float* r1 = start + (2 * h + 1) * inputWidth;
+        const float* r2 = start + (2 * h + 2) * inputWidth;
+        const float* r3 = start + (2 * h + 3) * inputWidth;
+        for (int s = 0; s < steps; s++) {
+          // Load the inputs
+          float32x4x2_t data1;
+          float32x4x2_t data2;
+
+          data1 = vld2q_f32(r0);
+          data2 = vld2q_f32(r0 + 8);
+          input[0][0] = data1.val[0];
+          input[0][1] = data1.val[1];
+          input[0][2] = vextq_f32(data1.val[0], data2.val[0], 1);
+          input[0][3] = vextq_f32(data1.val[1], data2.val[1], 1);
+
+          data1 = vld2q_f32(r1);
+          data2 = vld2q_f32(r1 + 8);
+          input[1][0] = data1.val[0];
+          input[1][1] = data1.val[1];
+          input[1][2] = vextq_f32(data1.val[0], data2.val[0], 1);
+          input[1][3] = vextq_f32(data1.val[1], data2.val[1], 1);
+
+          data1 = vld2q_f32(r2);
+          data2 = vld2q_f32(r2 + 8);
+          input[2][0] = data1.val[0];
+          input[2][1] = data1.val[1];
+          input[2][2] = vextq_f32(data1.val[0], data2.val[0], 1);
+          input[2][3] = vextq_f32(data1.val[1], data2.val[1], 1);
+
+          data1 = vld2q_f32(r3);
+          data2 = vld2q_f32(r3 + 8);
+          input[3][0] = data1.val[0];
+          input[3][1] = data1.val[1];
+          input[3][2] = vextq_f32(data1.val[0], data2.val[0], 1);
+          input[3][3] = vextq_f32(data1.val[1], data2.val[1], 1);
+
+          float32x4_t tmp1 = vdupq_n_f32(0.f);
+          float32x4_t tmp2 = vdupq_n_f32(0.f);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[0][0], k[0], 0);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[0][1], k[0], 1);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[0][2], k[0], 2);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[0][3], k[0], 3);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[1][0], k[1], 0);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[1][1], k[1], 1);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[1][2], k[1], 2);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[1][3], k[1], 3);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[2][0], k[2], 0);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[2][1], k[2], 1);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[2][2], k[2], 2);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[2][3], k[2], 3);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[3][0], k[3], 0);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[3][1], k[3], 1);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[3][2], k[3], 2);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[3][3], k[3], 3);
+          tmp1 = vaddq_f32(tmp1, tmp2);
+
+          vst1q_f32(outputData, tmp1);
+          r0 += 8;
+          r1 += 8;
+          r2 += 8;
+          r3 += 8;
+          outputData += 4;
+        }
+
+        for (int r = 0; r < remain; r++) {
+          float32x4_t i0 = vld1q_f32(r0);
+          float32x4_t i1 = vld1q_f32(r1);
+          float32x4_t i2 = vld1q_f32(r2);
+          float32x4_t i3 = vld1q_f32(r3);
+          *outputData = conv4x4(i0, i1, i2, i3, k[0], k[1], k[2], k[3]);
+          r0 += 2;
+          r1 += 2;
+          r2 += 2;
+          r3 += 2;
+          outputData++;
+        }
+      }
+    }
+  }
+};
+
+template <DeviceType Device>
+class NeonDepthwiseConvFunction : public ConvFunctionBase {
+public:
+  void init(const FuncConfig& config) override {
+    ConvFunctionBase::init(config);
+  }
+
+  void check(const BufferArgs& inputs, const BufferArgs& outputs) override {
+    const TensorShape& input = inputs[0].shape();
+    const TensorShape& filter = inputs[1].shape();
+    const TensorShape& output = outputs[0].shape();
+    checkShape(input, filter, output);
+  }
+
+  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
+    CHECK_EQ(numInputs_, inputs.size());
+    CHECK_EQ(numOutputs_, outputs.size());
+    check(inputs, outputs);
+
+    const TensorShape& input = inputs[0].shape();
+    const TensorShape& filter = inputs[1].shape();
+    const TensorShape& output = outputs[0].shape();
+
+    size_t batchSize = input[0];
+    size_t inputChannels = input[1];
+    size_t inputHeight = input[2];
+    size_t inputWidth = input[3];
+    size_t filterHeight = getFilterHeight(filter);
+    size_t filterWidth = getFilterWidth(filter);
+    size_t outputChannels = output[1];
+    size_t outputHeight = output[2];
+    size_t outputWidth = output[3];
+    size_t filterMultiplier = outputChannels / groups_;
+    CHECK_EQ(inputChannels, groups_);
+
+    // only support strideH() == strideW() and filterHeight == filterWidth.
+    CHECK_EQ(strideH(), strideW());
+    CHECK_EQ(filterHeight, filterWidth);
+
+    float* inputData = inputs[0].data<float>();
+    float* filterData = inputs[1].data<float>();
+    float* outputData = outputs[0].data<float>();
+
+    // padding the input
+    float* inputPadding = inputData;
+    if (paddingH() > 0 || paddingW() > 0) {
+      int newSize = batchSize * inputChannels * (inputHeight + 2 * paddingH()) *
+                    (inputWidth + 2 * paddingW());
+      resizeBuffer<Device>(newSize);
+      inputPadding = reinterpret_cast<float*>(memory_->getBuf());
+      Padding<float>::run(inputData,
+                          inputPadding,
+                          batchSize * inputChannels,
+                          inputHeight,
+                          inputWidth,
+                          paddingH(),
+                          paddingW());
+
+      // height and width of padding data
+      inputHeight += 2 * paddingH();
+      inputWidth += 2 * paddingW();
+    }
+
+    std::function<void(
+        const float*, const float*, int, int, int, int, int, int, float*)>
+        DepthWiseConv;
+
+    if (filterWidth == 3 && strideW() == 1) {
+      DepthWiseConv = DepthwiseConvKernel<3, 1>::run;
+    } else if (filterWidth == 3 && strideW() == 2) {
+      DepthWiseConv = DepthwiseConvKernel<3, 2>::run;
+    } else if (filterWidth == 4 && strideW() == 1) {
+      DepthWiseConv = DepthwiseConvKernel<4, 1>::run;
+    } else if (filterWidth == 4 && strideW() == 2) {
+      DepthWiseConv = DepthwiseConvKernel<4, 2>::run;
+    } else {
+      LOG(FATAL) << "Not supported";
+    }
+
+    for (size_t i = 0; i < batchSize; i++) {
+      DepthWiseConv(inputPadding,
+                    filterData,
+                    inputHeight,
+                    inputWidth,
+                    outputChannels,
+                    outputHeight,
+                    outputWidth,
+                    filterMultiplier,
+                    outputData);
+      inputPadding += inputChannels * inputHeight * inputWidth;
+      outputData += outputChannels * outputHeight * outputWidth;
+    }
+  }
+};
+
+REGISTER_TYPED_FUNC(NeonDepthwiseConv, CPU, NeonDepthwiseConvFunction);
+
+#endif
+
+}  // namespace neon
+}  // namespace paddle

paddle/function/neon/neon_util.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+
+#include <arm_neon.h>
+
+namespace paddle {
+
+namespace neon {
+
+inline float32x4_t vld1q_f32_aligned(const float* p) {
+  return vld1q_f32(
+      (const float*)__builtin_assume_aligned(p, sizeof(float32x4_t)));
+}
+
+#ifndef __aarch64__
+inline float32_t vaddvq_f32(float32x4_t a) {
+  float32x2_t v = vadd_f32(vget_high_f32(a), vget_low_f32(a));
+  return vget_lane_f32(vpadd_f32(v, v), 0);
+}
+
+inline float32x4_t vmlaq_laneq_f32(float32x4_t a,
+                                   float32x4_t b,
+                                   float32x4_t v,
+                                   const int lane) {
+  return vmlaq_n_f32(a, b, vgetq_lane_f32(v, lane));
+}
+#endif
+
+}  // namespace neon
+}  // namespace paddle
+
+#endif

paddle/gserver/layers/Conv3DLayer.cpp

+/* Copyright (c) 2016 Baidu, Inc. 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 "Conv3DLayer.h"
+#include "paddle/utils/Logging.h"
+#include "paddle/utils/Stat.h"
+
+namespace paddle {
+
+REGISTER_LAYER(conv3d, Conv3DLayer);
+
+bool Conv3DLayer::init(const LayerMap &layerMap,
+                       const ParameterMap &parameterMap) {
+  if (!ConvBaseLayer::init(layerMap, parameterMap)) return false;
+  int index = 0;
+  for (auto &inputConfig : config_.inputs()) {
+    const ConvConfig &conf = inputConfig.conv_conf();
+    M_.push_back(numFilters_ / conf.groups());
+    K_.push_back(filterPixels_[index] * filterChannels_[index]);
+
+    // create a new weight
+    size_t height, width;
+    width = filterPixels_[index] * filterChannels_[index];
+    height = numFilters_;
+    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_));
+    } else {
+      biases_ =
+          std::unique_ptr<Weight>(new Weight(1, getSize(), biasParameter_));
+    }
+  }
+  return true;
+}
+
+size_t Conv3DLayer::getSize() {
+  CHECK_NE(inputLayers_.size(), 0UL);
+  outputH_.clear();
+  outputW_.clear();
+  outputD_.clear();
+  N_.clear();
+  size_t layerSize = 0;
+  for (size_t i = 0; i < inputLayers_.size(); ++i) {
+    outputW_.push_back(outputSize(
+        imgSizeW_[i], filterSize_[i], padding_[i], stride_[i], true));
+    outputH_.push_back(outputSize(
+        imgSizeH_[i], filterSizeY_[i], paddingY_[i], strideY_[i], true));
+    outputD_.push_back(outputSize(
+        imgSizeD_[i], filterSizeZ_[i], paddingZ_[i], strideZ_[i], true));
+
+    N_.push_back(outputD_[i] * outputH_[i] * outputW_[i]);
+    CHECK(layerSize == 0 || N_[i] * size_t(numFilters_) == layerSize);
+    layerSize += N_[i] * numFilters_;
+  }
+  getOutput().setFrameHeight(outputH_[0]);
+  getOutput().setFrameWidth(outputW_[0]);
+  getOutput().setFrameDepth(outputD_[0]);
+  return layerSize;
+}
+
+void Conv3DLayer::forward(PassType passType) {
+  Layer::forward(passType);
+
+  int batchSize = inputLayers_[0]->getOutputValue()->getHeight();
+  int outWidth = getSize();
+  resetOutput(batchSize, outWidth);
+
+  for (size_t i = 0; i != inputLayers_.size(); ++i) {
+    REGISTER_TIMER_INFO("FwdConv3D", getName().c_str());
+    const MatrixPtr &inMat = getInputValue(i);
+    const MatrixPtr &outMat = getOutputValue();
+    int M = M_[i];
+    int N = N_[i];
+    int K = K_[i];
+    Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_);
+    MatrixPtr wMat = weights_[i]->getW();
+    for (int n = 0; n < batchSize; ++n) {
+      colBuf_->vol2Col(inMat->getData() + n * inMat->getStride(),
+                       channels_[i],
+                       imgSizeD_[i],
+                       imgSizeH_[i],
+                       imgSizeW_[i],
+                       filterSizeZ_[i],
+                       filterSizeY_[i],
+                       filterSize_[i],
+                       strideZ_[i],
+                       strideY_[i],
+                       stride_[i],
+                       paddingZ_[i],
+                       paddingY_[i],
+                       padding_[i]);
+
+      real *outData = outMat->getData() + n * outMat->getStride();
+      MatrixPtr outMatSub =
+          Matrix::create(outData, groups_[i] * M, N, false, useGpu_);
+      for (int g = 0; g < groups_[i]; g++) {
+        MatrixPtr wMatSub = wMat->subMatrix(g * M, M);
+        MatrixPtr in = colBuf_->subMatrix(g * K, K);
+        MatrixPtr out = outMatSub->subMatrix(g * M, M);
+        out->mul(*wMatSub, *in, 1.0, 1.0);
+      }
+    }
+  }
+  if (nullptr != this->biasParameter_) {
+    REGISTER_TIMER_INFO("FwBiasTimer", getName().c_str());
+    this->addBias();
+  }
+  forwardActivation();
+}
+
+void Conv3DLayer::backward(const UpdateCallback &callback) {
+  backwardActivation();
+
+  if (biases_ && biases_->getWGrad()) {
+    bpropBiases();
+    biases_->getParameterPtr()->incUpdate(callback);
+  }
+
+  for (size_t i = 0; i != inputLayers_.size(); ++i) {
+    REGISTER_TIMER_INFO("BwdConv3D", getName().c_str());
+    if (weights_[i]->getWGrad()) {
+      bpropWeights(i);
+    }
+    if (getInputGrad(i)) {
+      bpropData(i);
+    }
+    REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
+    weights_[i]->getParameterPtr()->incUpdate(callback);
+  }
+}
+
+void Conv3DLayer::bpropWeights(int i) {
+  int M = M_[i];
+  int N = N_[i];
+  int K = K_[i];
+  const MatrixPtr &inMat = getInputValue(i);
+  Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_);
+  MatrixPtr wGradMat = weights_[i]->getWGrad();
+  int batchSize = inputLayers_[0]->getOutputValue()->getHeight();
+  for (int n = 0; n < batchSize; ++n) {
+    colBuf_->vol2Col(inMat->getData() + n * inMat->getStride(),
+                     channels_[i],
+                     imgSizeD_[i],
+                     imgSizeH_[i],
+                     imgSizeW_[i],
+                     filterSizeZ_[i],
+                     filterSizeY_[i],
+                     filterSize_[i],
+                     strideZ_[i],
+                     strideY_[i],
+                     stride_[i],
+                     paddingZ_[i],
+                     paddingY_[i],
+                     padding_[i]);
+
+    real *outGradData =
+        getOutputGrad()->getData() + n * getOutputGrad()->getStride();
+    MatrixPtr outGradSub =
+        Matrix::create(outGradData, groups_[i] * M, N, false, useGpu_);
+    for (int g = 0; g < groups_[i]; ++g) {
+      MatrixPtr inMatSub = colBuf_->subMatrix(g * K, K);
+      MatrixPtr outG = outGradSub->subMatrix(g * M, M);
+      MatrixPtr wGradSub = wGradMat->subMatrix(g * M, M);
+      wGradSub->mul(*outG, *(inMatSub->getTranspose()), 1.0, 1.0);
+    }
+  }
+}
+
+void Conv3DLayer::bpropData(int i) {
+  int M = M_[i];
+  int N = N_[i];
+  int K = K_[i];
+  Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_);
+  MatrixPtr wMat = weights_[i]->getW();
+  int batchSize = inputLayers_[0]->getOutputValue()->getHeight();
+  for (int n = 0; n < batchSize; ++n) {
+    real *outGradData =
+        getOutputGrad()->getData() + n * getOutputGrad()->getStride();
+    real *preGradData =
+        getInputGrad(i)->getData() + n * getInputGrad(i)->getStride();
+    MatrixPtr outGradSub =
+        Matrix::create(outGradData, M * groups_[i], N, false, useGpu_);
+    for (int g = 0; g < groups_[i]; ++g) {
+      MatrixPtr wMatSub = wMat->subMatrix(g * M, M);
+      MatrixPtr outG = outGradSub->subMatrix(g * M, M);
+      MatrixPtr inGradMatSub = colBuf_->subMatrix(g * K, K);
+      inGradMatSub->mul(*(wMatSub->getTranspose()), *outG, 1.0, 0.0);
+    }
+    colBuf_->col2Vol(preGradData,
+                     channels_[i],
+                     imgSizeD_[i],
+                     imgSizeH_[i],
+                     imgSizeW_[i],
+                     filterSizeZ_[i],
+                     filterSizeY_[i],
+                     filterSize_[i],
+                     strideZ_[i],
+                     strideY_[i],
+                     stride_[i],
+                     paddingZ_[i],
+                     paddingY_[i],
+                     padding_[i],
+                     1.0,
+                     1.0);
+  }
+}
+
+void Conv3DLayer::bpropBiases() {
+  MatrixPtr outGradMat = getOutputGrad();
+  if (this->sharedBiases_) {
+    biases_->getWGrad()->collectSharedBias(*outGradMat, 1.0f);
+  } else {
+    biases_->getWGrad()->collectBias(*outGradMat, 1.0f);
+  }
+}
+
+void Conv3DLayer::addBias() {
+  MatrixPtr outMat = getOutputValue();
+  if (this->sharedBiases_) {
+    outMat->addSharedBias(*(biases_->getW()), 1.0f);
+  } else {
+    outMat->addBias(*(biases_->getW()), 1.0f);
+  }
+}
+
+}  // namespace paddle

paddle/gserver/layers/Conv3DLayer.h

+/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include <vector>
+#include "ConvBaseLayer.h"
+#include "paddle/math/MathUtils.h"
+#include "paddle/math/Matrix.h"
+
+namespace paddle {
+
+/**
+ * @brief A subclass of convolution layer.
+ * This layer expands input and use matrix multiplication to
+ * calculate convolution operation.
+ */
+class Conv3DLayer : public ConvBaseLayer {
+public:
+  explicit Conv3DLayer(const LayerConfig& config) : ConvBaseLayer(config) {}
+  ~Conv3DLayer() {}
+
+  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+
+  void forward(PassType passType);
+  void addBias();
+  void backward(const UpdateCallback& callback);
+  void bpropBiases();
+  void bpropData(int i);
+  void bpropWeights(int i);
+  size_t getSize();
+
+protected:
+  // Figure out the dimensions for individual gemms.
+  IntV M_;  /// numFilters_ / filter_group_;
+  IntV N_;  /// channels_ * filterSizeZ_ * filterSize_ * filterSizeY_
+  IntV K_;  /// outputD_ * outputH_ * outputW_
+  MatrixPtr colBuf_;
+};
+
+}  // namespace paddle

paddle/gserver/layers/ConvBaseLayer.cpp

@@ -38,7 +38,6 @@ bool ConvBaseLayer::init(const LayerMap& layerMap,
     strideY_.push_back(conf.stride_y());
     dilationY_.push_back(conf.dilation_y());
     filterSizeY_.push_back(conf.filter_size_y());
-    filterPixels_.push_back(filterSize_.back() * filterSizeY_.back());
     channels_.push_back(conf.channels());
     imgSizeH_.push_back(conf.has_img_size_y() ? conf.img_size_y()
                                               : conf.img_size());
@@ -47,31 +46,20 @@ bool ConvBaseLayer::init(const LayerMap& layerMap,
     filterChannels_.push_back(conf.filter_channels());
     outputH_.push_back(conf.has_output_y() ? conf.output_y() : conf.output_x());
     outputW_.push_back(conf.output_x());
+
+    paddingZ_.push_back(conf.padding_z());
+    strideZ_.push_back(conf.stride_z());
+    filterSizeZ_.push_back(conf.filter_size_z());
+    imgSizeD_.push_back(conf.img_size_z());
+    outputD_.push_back(conf.output_z());
+    filterPixels_.push_back(filterSize_.back() * filterSizeY_.back() *
+                            filterSizeZ_.back());
   }
 
   CHECK(inputLayers_.size() == parameters_.size());
-  for (size_t i = 0; i < inputLayers_.size(); i++) {
-    size_t height, width;
-    height = filterPixels_[i] * filterChannels_[i];
-    width = (!isDeconv_) ? numFilters_ : channels_[i];
-
-    // create a new weight
-    CHECK_EQ(parameters_[i]->getSize(), width * height);
-    Weight* w = new Weight(height, width, parameters_[i]);
-    weights_.emplace_back(w);
-  }
 
-  /* initialize the biases_ */
-  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_));
-    }
-  }
+  // create new weights_ in derived class
+  // create new biases_ in derived class
 
   // default caffe model
   caffeMode_ = true;

paddle/gserver/layers/ConvBaseLayer.h

@@ -62,6 +62,13 @@ protected:
   IntV outputH_;
   /// The spatial dimensions of output feature map width.
   IntV outputW_;
+
+  IntV outputD_;
+  IntV imgSizeD_;
+  IntV filterSizeZ_;
+  IntV strideZ_;
+  IntV paddingZ_;
+
   /// Group size, refer to grouped convolution in
   /// Alex Krizhevsky's paper: when group=2, the first half of the
   /// filters are only connected to the first half of the input channels,

paddle/gserver/layers/CostLayer.h

@@ -318,7 +318,9 @@ public:
 
   void forwardImp(Matrix& output, Argument& label, Matrix& cost) override;
 
-  void backwardImp(Matrix& outputValue, Argument& label, Matrix& outputGrad) {}
+  void backwardImp(Matrix& outputValue,
+                   Argument& label,
+                   Matrix& outputGrad) override {}
 };
 
 /**

paddle/gserver/layers/CrossEntropyOverBeam.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 "CrossEntropyOverBeam.h"
+
+namespace paddle {
+
+void CostForOneSequence::calValidExpandStep() {
+  validExpansionCount_ = 0;
+  goldAsExtraPath_ = true;
+
+  for (size_t i = 0; i < beams_->expansionCount; ++i) {
+    real gold = static_cast<real>(beams_->gold[i]);
+    if (i) {
+      real* start = beams_->candidateIds[i - 1]->getData();
+      goldRowIds_[i] = std::count_if(
+          start,
+          start + goldRowIds_[i - 1] * beamSize_ + goldColIds_[i - 1],
+          [](const real& val) { return val != -1.; });
+    } else {
+      goldRowIds_[i] = 0;
+    }
+
+    real* start =
+        beams_->candidateIds[i]->getData() + goldRowIds_[i] * beamSize_;
+    real* findEnd = std::find(start, start + beamSize_, gold);
+    validExpansionCount_++;
+
+    if (start + beamSize_ == findEnd) return;
+    goldColIds_[i] = findEnd - start;
+  }
+  if (goldColIds_[beams_->expansionCount - 1] != -1) goldAsExtraPath_ = false;
+}
+
+size_t CostForOneSequence::initLastExpansion() {
+  int beamId = validExpansionCount_ - 1;
+  const MatrixPtr candidates = beams_->candidateIds[beamId];
+  size_t height = candidates->getHeight();
+
+  /* initialization the last expansion. */
+  size_t pathCount = std::count_if(candidates->getData(),
+                                   candidates->getData() + height * beamSize_,
+                                   [](const real& val) { return val != -1; });
+  /*
+   * if the gold sequence falls off the beam during search, add the gold
+   * sequence as the last path into the all expanded candidates.
+   */
+  if (goldAsExtraPath_) goldIdsInFinalExpansion_ = pathCount++;
+
+  pathRowIdsInEachBeam_.clear();
+  pathRowIdsInEachBeam_.resize(validExpansionCount_,
+                               std::vector<int>(pathCount, 0));
+  parentIdsInBeam_.clear();
+  parentIdsInBeam_.resize(pathCount, 0);
+
+  if (goldAsExtraPath_) {
+    /* add gold sequence into the total expansion. */
+    pathRowIdsInEachBeam_[beamId].back() =
+        beams_->gold[beamId] +
+        getSeqStartPos(beamId, goldRowIds_[validExpansionCount_ - 1]);
+    parentIdsInBeam_.back() = goldRowIds_[validExpansionCount_ - 1];
+  } else {
+    size_t goldOffset = goldRowIds_[beamId] * beamSize_ + goldColIds_[beamId];
+    goldIdsInFinalExpansion_ =
+        std::count_if(candidates->getData(),
+                      candidates->getData() + goldOffset,
+                      [](const real& val) { return val != -1.; });
+  }
+
+  /*
+   * TODO(caoying): fix this, store the indices of selected candidate
+   * paths into Argument.ids
+   */
+  real* ids = candidates->getData();
+  size_t curIdx = 0;
+  for (size_t i = 0; i < height; ++i) {
+    int basePos = getSeqStartPos(beamId, i);
+    for (size_t j = 0; j < beamSize_; ++j) {
+      int id = ids[i * beamSize_ + j];
+      if (id == -1) continue;
+      pathRowIdsInEachBeam_[beamId][curIdx] = id + basePos;
+      parentIdsInBeam_[curIdx++] = i;
+    }
+  }
+  return pathCount;
+}
+
+void CostForOneSequence::constructTotalExpansion() {
+  /*
+   * construct the entire expanded beam by begining with the last search
+   * in which gold falls off the beam.
+   */
+  size_t totalPathCount = initLastExpansion();
+
+  for (int beamId = validExpansionCount_ - 2; beamId >= 0; --beamId) {
+    const MatrixPtr candidates = beams_->candidateIds[beamId];
+    real* ids = candidates->getData();
+
+    int lastParentIdInBeam = -1;
+    int basePos = -1;
+    for (size_t i = 0;
+         i < (goldAsExtraPath_ ? totalPathCount - 1 : totalPathCount);
+         ++i) {
+      int id = ids[parentIdsInBeam_[i]];
+      int parentRowId = std::div(parentIdsInBeam_[i], beamSize_).quot;
+      if (parentIdsInBeam_[i] != lastParentIdInBeam)
+        basePos = getSeqStartPos(beamId, parentRowId);
+
+      pathRowIdsInEachBeam_[beamId][i] = id + basePos;
+      lastParentIdInBeam = parentIdsInBeam_[i];
+      parentIdsInBeam_[i] = parentRowId;
+
+      if (goldAsExtraPath_)
+        pathRowIdsInEachBeam_[beamId][totalPathCount - 1] =
+            beams_->gold[beamId] + getSeqStartPos(beamId, goldRowIds_[beamId]);
+    }
+  }
+}
+
+real CostForOneSequence::globallyNormalizedScore() {
+  expandedPathScores_.resize(validExpansionCount_);
+
+  Matrix::resizeOrCreate(
+      softmaxOut_, 1, pathRowIdsInEachBeam_[0].size(), false, false);
+  softmaxOut_->zeroMem();
+  MatrixPtr tmp = Matrix::create(
+      softmaxOut_->getData(), softmaxOut_->getWidth(), 1, false, false);
+
+  for (size_t i = 0; i < validExpansionCount_; ++i) {
+    Matrix::resizeOrCreate(expandedPathScores_[i],
+                           pathRowIdsInEachBeam_[i].size(),
+                           1,
+                           false,
+                           false);
+    expandedPathScores_[i]->zeroMem();
+
+    IVectorPtr rowIds = IVector::create(pathRowIdsInEachBeam_[i].data(),
+                                        pathRowIdsInEachBeam_[i].size(),
+                                        false);
+    expandedPathScores_[i]->selectRows(*(beams_->scores[i]), *rowIds);
+    tmp->add(*expandedPathScores_[i]);
+  }
+
+  softmaxOut_->softmax(*softmaxOut_);
+  return -std::log(softmaxOut_->getData()[goldIdsInFinalExpansion_]);
+}
+
+real CostForOneSequence::forward() {
+  calValidExpandStep();
+  constructTotalExpansion();
+  return globallyNormalizedScore();
+}
+
+void CostForOneSequence::backward() {
+  /*
+   * when softmax layer is the output layer, and it is combined with
+   * cross-entropy as cost. The derivate with regard to softmax's input
+   * is simply:
+   *
+   * grad_i = softmax_out_i - target_i,
+   *
+   * and here hard label is used.
+   */
+  softmaxOut_->getData()[goldIdsInFinalExpansion_] -= 1.;
+
+  MatrixPtr tmp = Matrix::create(
+      softmaxOut_->getData(), softmaxOut_->getWidth(), 1, false, false);
+
+  for (size_t i = 0; i < validExpansionCount_; ++i) {
+    IVectorPtr rowIds = IVector::create(pathRowIdsInEachBeam_[i].data(),
+                                        pathRowIdsInEachBeam_[i].size(),
+                                        false);
+    /*
+      beams_->scoreGrad[i] has been intialized outside this class, this
+      class only keeps a pointer pointing to the original input gradients,
+      so here does not need to allocate or initalize the memory.
+    */
+    tmp->addToRows(*beams_->scoreGrad[i], *rowIds);
+  }
+}
+
+REGISTER_LAYER(cross_entropy_over_beam, CrossEntropyOverBeam);
+
+bool CrossEntropyOverBeam::init(const LayerMap& layerMap,
+                                const ParameterMap& parameterMap) {
+  /* Initialize the basic parent class */
+  Layer::init(layerMap, parameterMap);
+  CHECK_EQ(0U, inputLayers_.size() % 3) << "Error input number.";
+
+  beamExpanCount_ = inputLayers_.size() / 3;
+
+  candidateScores_.resize(beamExpanCount_);
+  candidateScoreGrad_.resize(beamExpanCount_);
+
+  candidateInBeam_.resize(beamExpanCount_);
+  goldSequence_.resize(beamExpanCount_);
+  gradToInputs_.resize(beamExpanCount_);
+
+  setNeedSequenceInfo(false);
+  return true;
+}
+
+void CrossEntropyOverBeam::checkInputs() {
+  batchSize_ = 0;
+  for (size_t i = 0; i < beamExpanCount_; ++i) {
+    const Argument& scores = getInput(i * 3);
+    const Argument& selCandidates = getInput(i * 3 + 1);
+    const Argument& goldSeq = getInput(i * 3 + 2);
+
+    if (i) {
+      CHECK(scores.hasSubseq()) << "input " << i << " "
+                                << inputLayers_[i * 3]->getName()
+                                << " should be a nested sequence";
+      CHECK_EQ(getInputValue(i * 3 + 1)->getWidth(), beamSize_);
+      CHECK_EQ(batchSize_, static_cast<size_t>(scores.getNumSequences()));
+      CHECK_EQ(scores.getNumSubSequences(), selCandidates.getBatchSize());
+    } else {
+      CHECK(scores.hasSeq()) << "input " << i << " "
+                             << inputLayers_[i]->getName()
+                             << " should be a sequence";
+      batchSize_ = scores.getNumSequences();
+      beamSize_ = getInputValue(i * 3 + 1)->getWidth();
+      CHECK_EQ(batchSize_, static_cast<size_t>(selCandidates.getBatchSize()));
+    }
+    CHECK_EQ(1U, scores.value->getWidth());
+    CHECK_EQ(batchSize_, static_cast<size_t>(goldSeq.getBatchSize()));
+  }
+}
+
+void CrossEntropyOverBeam::copyInputsToCpu() {
+  auto copyValue = [](const MatrixPtr& src, MatrixPtr& trg) {
+    if (dynamic_cast<GpuMatrix*>(src.get())) {
+      Matrix::resizeOrCreate(
+          trg, src->getHeight(), src->getWidth(), false, false);
+      trg->copyFrom(*src);
+    } else {
+      trg = std::move(src);
+    }
+  };
+
+  auto copyIds = [](const IVectorPtr& src, IVectorPtr& trg) {
+    if (dynamic_cast<GpuIVector*>(src.get())) {
+      IVector::resizeOrCreate(trg, src->getSize(), false);
+      trg->copyFrom(*src);
+    } else {
+      trg = std::move(src);
+    }
+  };
+
+  beamSplitPos_.clear();
+  beamSplitPos_.resize(batchSize_, std::vector<int>(beamExpanCount_, 0));
+  for (size_t i = 0; i < beamExpanCount_; ++i) {
+    copyValue(getInputValue(i * 3), candidateScores_[i]);
+    copyValue(getInputValue(i * 3 + 1), candidateInBeam_[i]);
+    copyIds(getInput(i * 3 + 2).ids, goldSequence_[i]);
+
+    if (i) {
+      ICpuGpuVectorPtr seqInfo = getInput(i * 3).sequenceStartPositions;
+      const int* seqStarts = seqInfo->getMutableData(false);
+      ICpuGpuVectorPtr subSeqInfo = getInput(i * 3).subSequenceStartPositions;
+      const int* subSeqStarts = subSeqInfo->getMutableData(false);
+
+      size_t seqId = 1;
+      for (size_t subSeqId = 0; subSeqId < subSeqInfo->getSize() - 1;
+           ++subSeqId) {
+        CHECK_LT(seqId, seqInfo->getSize());
+        if (subSeqStarts[subSeqId] == seqStarts[seqId]) {
+          beamSplitPos_[seqId][i] = beamSplitPos_[seqId - 1][i];
+          seqId++;
+        }
+        beamSplitPos_[seqId - 1][i]++;
+      }
+    } else {
+      for (size_t j = 0; j < batchSize_; ++j) beamSplitPos_[j][i] = j + 1;
+    }
+  }
+}
+
+void CrossEntropyOverBeam::splitBatchBeams() {
+  beamCosts_.resize(batchSize_);
+  beamPerSeq_.resize(batchSize_, BeamExpansion(beamExpanCount_));
+
+  for (size_t i = 0; i < beamExpanCount_; ++i) {
+    int* seqStarts =
+        getInput(i * 3).sequenceStartPositions->getMutableData(false);
+
+    int* subSeqStarts = nullptr;
+    int maxLen = 0;
+    if (i) {
+      subSeqStarts =
+          getInput(i * 3).subSequenceStartPositions->getMutableData(false);
+      maxLen = getInput(i * 3).subSequenceStartPositions->getSize() - 1;
+    } else {
+      maxLen = getInput(i).sequenceStartPositions->getSize() - 1;
+    }
+
+    for (size_t j = 0; j < batchSize_; ++j) {
+      beamPerSeq_[j].scores[i] =
+          Matrix::create(candidateScores_[i]->getData() + seqStarts[j],
+                         seqStarts[j + 1] - seqStarts[j],
+                         1,
+                         false,
+                         false);
+      beamPerSeq_[j].scoreGrad[i] =
+          Matrix::create(candidateScoreGrad_[i]->getData() + seqStarts[j],
+                         seqStarts[j + 1] - seqStarts[j],
+                         1,
+                         false,
+                         false);
+
+      int offset = j ? beamSplitPos_[j - 1][i] : 0;
+      int height = beamSplitPos_[j][i] - (j ? beamSplitPos_[j - 1][i] : 0);
+      CHECK_GE(maxLen, offset + height);
+      beamPerSeq_[j].seqInfo[i] = IVector::create(
+          (i ? subSeqStarts : seqStarts) + offset, height + 1, false);
+
+      beamPerSeq_[j].candidateIds[i] =
+          Matrix::create(candidateInBeam_[i]->getData() + offset * beamSize_,
+                         height,
+                         beamSize_,
+                         false,
+                         false);
+      beamPerSeq_[j].gold[i] = goldSequence_[i]->getData()[j];
+
+      CHECK_LE(beamPerSeq_[j].gold[i], seqStarts[j + 1] - seqStarts[j]);
+    }
+  }
+}
+
+void CrossEntropyOverBeam::resizeOutput() {
+  Matrix::resizeOrCreate(output_.value, batchSize_, 1, false, false);
+  output_.value->zeroMem();
+
+  for (size_t i = 0; i < beamExpanCount_; ++i) {
+    MatrixPtr inGrad = getInputGrad(i * 3);
+    if (dynamic_cast<GpuMatrix*>(inGrad.get())) {
+      Matrix::resizeOrCreate(candidateScoreGrad_[i],
+                             inGrad->getHeight(),
+                             inGrad->getWidth(),
+                             false,
+                             false);
+    } else {
+      candidateScoreGrad_[i] = std::move(inGrad);
+    }
+    candidateScoreGrad_[i]->zeroMem();
+  }
+}
+
+void CrossEntropyOverBeam::copyGradToGpu(size_t copyCount) {
+  for (size_t i = 0; i < beamExpanCount_; ++i) {
+    if (dynamic_cast<GpuMatrix*>(getInputGrad(i * 3).get()))
+      getInputGrad(i * 3)->copyFrom(*candidateScoreGrad_[i]);
+
+    if (i == copyCount - 1) break;
+  }
+}
+
+void CrossEntropyOverBeam::forward(PassType passType) {
+  Layer::forward(passType);
+
+  checkInputs();
+  copyInputsToCpu();
+
+  resizeOutput();
+  splitBatchBeams();
+
+  MatrixPtr outputValue = getOutputValue();
+  for (size_t i = 0; i < batchSize_; ++i) {
+    BeamExpansionPtr ptr = std::make_shared<BeamExpansion>(beamPerSeq_[i]);
+    beamCosts_[i].setData(std::move(ptr), beamSize_);
+    outputValue->getData()[i] = beamCosts_[i].forward();
+  }
+}
+
+void CrossEntropyOverBeam::backward(const UpdateCallback& callback) {
+  for (size_t i = 0; i < batchSize_; ++i) {
+    beamCosts_[i].backward();
+    copyGradToGpu(beamCosts_[i].getValidExpansionCount());
+  }
+}
+
+}  // namespace paddle

paddle/gserver/layers/CrossEntropyOverBeam.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 "CrossEntropyOverBeam.h"
+#include "Layer.h"
+
+namespace paddle {
+
+/* This struct stores the beams in all search steps for a single sequence. */
+struct BeamExpansion {
+  std::vector<MatrixPtr> scores;
+  std::vector<IVectorPtr> seqInfo;
+
+  std::vector<MatrixPtr> candidateIds;
+  std::vector<int> gold;
+
+  std::vector<MatrixPtr> scoreGrad;
+
+  size_t expansionCount;
+
+  explicit BeamExpansion(int n) {
+    expansionCount = n;
+    scores.resize(expansionCount);
+    seqInfo.resize(expansionCount);
+    candidateIds.resize(expansionCount);
+    scoreGrad.resize(expansionCount);
+
+    gold.resize(expansionCount);
+  }
+};
+typedef std::shared_ptr<BeamExpansion> BeamExpansionPtr;
+
+class CostForOneSequence {
+public:
+  CostForOneSequence()
+      : beamSize_(0), validExpansionCount_(0), goldAsExtraPath_(false) {}
+  void setData(const BeamExpansionPtr bPtr, size_t beamSize) {
+    beams_ = bPtr;
+    beamSize_ = beamSize;
+
+    expandedPathScores_.clear();
+    expandedPathScores_.resize(beams_->expansionCount);
+
+    goldRowIds_.clear();
+    goldRowIds_.resize(beams_->expansionCount, 0);
+    goldColIds_.clear();
+    goldColIds_.resize(beams_->expansionCount, -1);
+  }
+  size_t getValidExpansionCount() { return validExpansionCount_; }
+
+  real forward();
+  void backward();
+
+private:
+  void calValidExpandStep();
+  void constructTotalExpansion();
+  size_t initLastExpansion();
+  real globallyNormalizedScore();
+
+  int getSeqStartPos(size_t beamId, size_t rowId) {
+    CHECK_GT(beams_->seqInfo[beamId]->getSize() - 1, rowId);
+    int* starts = beams_->seqInfo[beamId]->getData();
+    return starts[rowId] - starts[0];
+  }
+
+  size_t beamSize_;
+  size_t validExpansionCount_;
+  bool goldAsExtraPath_;
+  std::vector<int> goldRowIds_;
+  std::vector<int> goldColIds_;
+
+  BeamExpansionPtr beams_;
+  std::vector<std::vector<int>> pathRowIdsInEachBeam_;
+  std::vector<int> parentIdsInBeam_;
+  size_t goldIdsInFinalExpansion_;
+
+  std::vector<MatrixPtr> expandedPathScores_;
+
+  MatrixPtr softmaxOut_;
+};
+
+class CrossEntropyOverBeam : public Layer {
+public:
+  explicit CrossEntropyOverBeam(const LayerConfig& config) : Layer(config) {}
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback) override;
+
+private:
+  void checkInputs();
+  void copyInputsToCpu();
+  void resizeOutput();
+  void copyGradToGpu(size_t copyCount);
+  void splitBatchBeams();
+
+  size_t beamExpanCount_;
+  size_t batchSize_;
+  size_t beamSize_;
+
+  /*
+   * the process of constructing beams is not friendly to GPU, currently, this
+   * layer only runs on CPU, if any of its inputs is on GPU memory, then copy
+   * it to CPU memory.
+   */
+  std::vector<MatrixPtr> candidateScores_;
+  std::vector<MatrixPtr> candidateScoreGrad_;
+  std::vector<MatrixPtr> candidateInBeam_;
+  std::vector<MatrixPtr> gradToInputs_;
+  std::vector<IVectorPtr> goldSequence_;
+  std::vector<std::vector<int>> beamSplitPos_;
+
+  /*
+   * split entire bath of beams into beam per sequnence and store the result
+   * into this member.
+   */
+  std::vector<BeamExpansion> beamPerSeq_;
+  /* beamCosts_ is used to propagate error in one sequence. */
+  std::vector<CostForOneSequence> beamCosts_;
+};
+
+}  // namespace paddle

paddle/gserver/layers/CudnnConvBaseLayer.cpp

@@ -46,8 +46,26 @@ bool CudnnConvBaseLayer::init(const LayerMap &layerMap,
     projConf_.emplace_back(conf);
     projections_.emplace_back(
         Projection::create(*projConf_[i], parameters_[i], useGpu_));
+
+    // create a new weight
+    size_t height, width;
+    height = filterPixels_[i] * filterChannels_[i];
+    width = (!isDeconv_) ? numFilters_ : channels_[i];
+    CHECK_EQ(parameters_[i]->getSize(), width * height);
+    Weight *w = new Weight(height, width, parameters_[i]);
+    weights_.emplace_back(w);
   }
 
+  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_));
+    }
+  }
   if (biases_.get() && sharedBiases_) {
     hl_create_tensor_descriptor(&biasDesc_);
     hl_create_tensor_descriptor(&outputDesc_);

paddle/gserver/layers/DeConv3DLayer.cpp

+/* Copyright (c) 2016 Baidu, Inc. 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 "DeConv3DLayer.h"
+#include "paddle/utils/Logging.h"
+#include "paddle/utils/Stat.h"
+
+namespace paddle {
+
+REGISTER_LAYER(deconv3d, DeConv3DLayer);
+
+bool DeConv3DLayer::init(const LayerMap &layerMap,
+                         const ParameterMap &parameterMap) {
+  if (!ConvBaseLayer::init(layerMap, parameterMap)) return false;
+  // for Deconv, the dimension of Kernel is
+  // channel * output * depth * height * weigth
+  // Matrix storage format: (output * depth * height * weigth) x  channel
+  for (int index = 0; index < config_.inputs().size(); ++index) {
+    M_.push_back(filterChannels_[index]);
+    K_.push_back(filterPixels_[index] * (numFilters_ / groups_[index]));
+
+    // create a new weight
+    size_t height, width;
+    height = filterPixels_[index] * numFilters_;
+    width = filterChannels_[index];
+    CHECK_EQ(parameters_[index]->getSize(), width * height);
+    Weight *w = new Weight(height, width, parameters_[index]);
+    weights_.emplace_back(w);
+  }
+  if (biasParameter_.get()) {
+    if (sharedBiases_) {
+      CHECK_EQ((size_t)numFilters_, biasParameter_->getSize());
+      biases_ =
+          std::unique_ptr<Weight>(new Weight(1, numFilters_, biasParameter_));
+    } else {
+      biases_ =
+          std::unique_ptr<Weight>(new Weight(1, getSize(), biasParameter_));
+    }
+  }
+  return true;
+}
+
+size_t DeConv3DLayer::getSize() {
+  CHECK_NE(inputLayers_.size(), 0UL);
+  outputH_.clear();
+  outputW_.clear();
+  outputD_.clear();
+  N_.clear();
+  NOut_.clear();
+  size_t layerSize = 0;
+  for (size_t i = 0; i < inputLayers_.size(); ++i) {
+    outputW_.push_back(
+        imageSize(imgSizeW_[i], filterSize_[i], padding_[i], stride_[i], true));
+    outputH_.push_back(imageSize(
+        imgSizeH_[i], filterSizeY_[i], paddingY_[i], strideY_[i], true));
+    outputD_.push_back(imageSize(
+        imgSizeD_[i], filterSizeZ_[i], paddingZ_[i], strideZ_[i], true));
+    NOut_.push_back(outputD_[i] * outputH_[i] * outputW_[i]);
+    N_.push_back(imgSizeD_[i] * imgSizeH_[i] * imgSizeW_[i]);
+    CHECK(layerSize == 0 || N_[i] * size_t(numFilters_) == layerSize);
+    layerSize += NOut_[i] * numFilters_;
+  }
+  getOutput().setFrameHeight(outputH_[0]);
+  getOutput().setFrameWidth(outputW_[0]);
+  getOutput().setFrameDepth(outputD_[0]);
+  return layerSize;
+}
+
+void DeConv3DLayer::forward(PassType passType) {
+  Layer::forward(passType);
+  int batchSize = inputLayers_[0]->getOutputValue()->getHeight();
+  int outWidth = getSize();
+  resetOutput(batchSize, outWidth);
+  const MatrixPtr outMat = getOutputValue();
+
+  for (size_t i = 0; i != inputLayers_.size(); ++i) {
+    REGISTER_TIMER_INFO("FwdDeConv3D", getName().c_str());
+    const MatrixPtr &inMat = getInputValue(i);
+    int M = M_[i];
+    int N = N_[i];
+    int K = K_[i];
+    MatrixPtr wMat = weights_[i]->getW();
+    Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_);
+    for (int n = 0; n < batchSize; ++n) {
+      real *inData = inMat->getData() + n * inMat->getStride();
+      for (int g = 0; g < groups_[i]; ++g) {
+        MatrixPtr inMatSub = Matrix::create(inData, M, N, false, useGpu_);
+        MatrixPtr wMatSub = wMat->subMatrix(g * K, K);
+        MatrixPtr colBufDataSub = colBuf_->subMatrix(g * K, K);
+        colBufDataSub->mul(*wMatSub, *inMatSub, 1.0, 0.0);
+        inData += M * N;
+      }
+      colBuf_->col2Vol(outMat->getData() + n * outMat->getStride(),
+                       numFilters_,
+                       outputD_[i],
+                       outputH_[i],
+                       outputW_[i],
+                       filterSizeZ_[i],
+                       filterSizeY_[i],
+                       filterSize_[i],
+                       strideZ_[i],
+                       strideY_[i],
+                       stride_[i],
+                       paddingZ_[i],
+                       paddingY_[i],
+                       padding_[i],
+                       1.0,
+                       1.0);
+    }
+  }
+  if (nullptr != this->biasParameter_) {
+    REGISTER_TIMER_INFO("FwBiasTimer", getName().c_str());
+    this->addBias();
+  }
+  forwardActivation();
+}
+
+void DeConv3DLayer::backward(const UpdateCallback &callback) {
+  backwardActivation();
+  int batchSize = getOutputGrad()->getHeight();
+  if (biases_ && biases_->getWGrad()) {
+    bpropBiases();
+    biases_->getParameterPtr()->incUpdate(callback);
+  }
+  for (size_t i = 0; i < inputLayers_.size(); ++i) {
+    if (weights_[i]->getWGrad() || this->needGradient_) {
+      int M = M_[i];
+      int N = N_[i];
+      int K = K_[i];
+      REGISTER_TIMER_INFO("BwdDeConv3D", getName().c_str());
+      Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_);
+      const MatrixPtr &inMat = getInputValue(i);
+      for (int n = 0; n < batchSize; ++n) {
+        colBuf_->vol2Col(
+            getOutputGrad()->getData() + n * getOutputGrad()->getStride(),
+            numFilters_,
+            outputD_[i],
+            outputH_[i],
+            outputW_[i],
+            filterSizeZ_[i],
+            filterSizeY_[i],
+            filterSize_[i],
+            strideZ_[i],
+            strideY_[i],
+            stride_[i],
+            paddingZ_[i],
+            paddingY_[i],
+            padding_[i]);
+        if (weights_[i]->getWGrad()) {
+          real *inData = inMat->getData() + n * inMat->getStride();
+          for (int g = 0; g < groups_[i]; ++g) {
+            MatrixPtr colBufDataSub = colBuf_->subMatrix(g * K, K);
+            MatrixPtr wGradMatSub =
+                weights_[i]->getWGrad()->subMatrix(g * K, K);
+            MatrixPtr inMatSub = Matrix::create(inData, M, N, false, useGpu_);
+            wGradMatSub->mul(
+                *colBufDataSub, *(inMatSub->getTranspose()), 1.0, 1.0);
+            inData += M * N;
+          }
+        }
+        if (getInputGrad(i)) {
+          real *preGrad =
+              getInputGrad(i)->getData() + n * getInputGrad(i)->getStride();
+          for (int g = 0; g < groups_[i]; ++g) {
+            MatrixPtr w = weights_[i]->getW()->subMatrix(g * K, K);
+            MatrixPtr outGradMat = colBuf_->subMatrix(g * K, K);
+            MatrixPtr inGradMatSub =
+                Matrix::create(preGrad, M, N, false, useGpu_);
+            inGradMatSub->mul(*(w->getTranspose()), *outGradMat, 1.0, 1.0);
+            preGrad += M * N;
+          }
+        }
+      }
+      REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
+      weights_[i]->getParameterPtr()->incUpdate(callback);
+    }
+  }
+}
+void DeConv3DLayer::bpropWeights(int i) {}
+void DeConv3DLayer::bpropData(int i) {}
+
+void DeConv3DLayer::bpropBiases() {
+  const MatrixPtr &outGradMat = getOutputGrad();
+
+  if (this->sharedBiases_) {
+    biases_->getWGrad()->collectSharedBias(*outGradMat, 1.0f);
+  } else {
+    biases_->getWGrad()->collectBias(*outGradMat, 1.0f);
+  }
+}
+
+void DeConv3DLayer::addBias() {
+  MatrixPtr outMat = getOutputValue();
+  if (this->sharedBiases_) {
+    outMat->addSharedBias(*(biases_->getW()), 1.0f);
+  } else {
+    outMat->addBias(*(biases_->getW()), 1.0f);
+  }
+}
+
+}  // namespace paddle

paddle/gserver/layers/DeConv3DLayer.h

+/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include <vector>
+#include "ConvBaseLayer.h"
+#include "paddle/math/MathUtils.h"
+#include "paddle/math/Matrix.h"
+
+namespace paddle {
+
+/**
+ * @brief A subclass of deconvolution3D layer.
+ * This layer expands input and use matrix multiplication to
+ * calculate deconvolution3D operation.
+ */
+class DeConv3DLayer : public ConvBaseLayer {
+public:
+  explicit DeConv3DLayer(const LayerConfig& config) : ConvBaseLayer(config) {}
+  ~DeConv3DLayer() {}
+  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+
+  void forward(PassType passType);
+  void addBias();
+  void backward(const UpdateCallback& callback);
+  void bpropBiases();
+  void bpropData(int i);
+  void bpropWeights(int i);
+  size_t getSize();
+
+protected:
+  // Figure out the dimensions for individual gemms.
+  IntV M_;  /// numFilters_ / filter_group_;
+  IntV N_;  /// channels_ * filterSizeZ_ * filterSize_ * filterSizeY_
+  IntV K_;  /// outputD_ * outputH_ * outputW_
+  IntV NOut_;
+  MatrixPtr colBuf_;
+};
+
+}  // namespace paddle

paddle/gserver/layers/ExpandConvBaseLayer.cpp

@@ -22,12 +22,31 @@ bool ExpandConvBaseLayer::init(const LayerMap &layerMap,
   /* 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;

paddle/gserver/layers/ExpandConvLayer.cpp

@@ -29,6 +29,10 @@ namespace paddle {
 REGISTER_LAYER(exconv, ExpandConvLayer);
 REGISTER_LAYER(exconvt, ExpandConvLayer);
 
+inline bool isDepthwiseConv(int channels, int groups) {
+  return channels == groups;
+}
+
 bool ExpandConvLayer::init(const LayerMap &layerMap,
                            const ParameterMap &parameterMap) {
   /* Initialize the basic convolutional parent class */
@@ -47,14 +51,27 @@ bool ExpandConvLayer::init(const LayerMap &layerMap,
     std::vector<size_t> paddings = {(size_t)paddingY_[i], (size_t)padding_[i]};
     std::vector<size_t> strides = {(size_t)strideY_[i], (size_t)stride_[i]};
 
-    if (useGpu_ && (size_t)groups_[i] == (size_t)channels_[i] && !isDeconv_) {
+    // Convolution Layer uses the GemmConv function by default.
+    convType = "GemmConv";
+    convGradInputType = "GemmConvGradInput";
+    convGradFilterType = "GemmConvGradFilter";
+
+    // If depth wise convolution and useGpu == true
+    if (useGpu_ && isDepthwiseConv(channels_[i], groups_[i]) && !isDeconv_) {
       convType = "DepthwiseConv";
       convGradInputType = "DepthwiseConvGradInput";
       convGradFilterType = "DepthwiseConvGradFilter";
-    } else {
-      convType = "GemmConv";
-      convGradInputType = "GemmConvGradInput";
-      convGradFilterType = "GemmConvGradFilter";
+    }
+
+    // If depth wise convolution and useGpu == false and ARM-NEON
+    if (!useGpu_ && isDepthwiseConv(channels_[i], groups_[i]) && !isDeconv_) {
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+      if ((filterSize_[i] == filterSizeY_[i]) &&
+          (filterSize_[i] == 3 || filterSize_[i] == 4) &&
+          (stride_[i] == strideY_[i]) && (stride_[i] == 1 || stride_[i] == 2)) {
+        convType = "NeonDepthwiseConv";
+      }
+#endif
     }
 
     if (FLAGS_use_nnpack && !isDeconv_) {

paddle/gserver/layers/Layer.cpp

@@ -41,7 +41,7 @@ namespace paddle {
 Layer::Layer(const LayerConfig& config, bool useGpu)
     : config_(config),
       useGpu_(useGpu),
-      deviceId_(-1),
+      deviceId_(CPU_DEVICE),
       needSequenceInfo_(true) {}
 
 bool Layer::init(const LayerMap& layerMap, const ParameterMap& parameterMap) {

paddle/gserver/layers/Layer.h

@@ -59,7 +59,12 @@ protected:
   LayerConfig config_;
   /// whether to use GPU
   bool useGpu_;
-  /// Device Id. CPU is -1, and GPU is 0, 1, 2 ...
+  /// Paddle device ID, MKLDNN is -2, CPU is -1
+  enum PADDLE_DEVICE_ID {
+    MKLDNN_DEVICE = -2,
+    CPU_DEVICE = -1,
+  };
+  /// Device Id. MKLDNN is -2, CPU is -1, and GPU is 0, 1, 2 ...
   int deviceId_;
   /// Input layers
   std::vector<LayerPtr> inputLayers_;
@@ -77,6 +82,7 @@ protected:
   Argument output_;
   /// Several outputs stored on different devices, used in 'parallel_nn' case,
   /// and record them by deviceId_.
+  /// Also used in 'use_mkldnn' case.
   std::vector<Argument> outputOtherDevice_;
   /// If there are several outputs, map them by each name.
   std::map<std::string, Argument*> outputMap_;
@@ -172,6 +178,13 @@ protected:
     return inputLayer.getOutput(deviceId_);
   }
 
+  /**
+   * Get the argument of input layer with deviceId.
+   */
+  const Argument& getInput(size_t inputIndex, int deviceId) const {
+    return inputLayers_[inputIndex]->getOutput(deviceId);
+  }
+
   /**
    * Get the forward-input value.
    */
@@ -186,6 +199,13 @@ protected:
     return inputLayer.getOutput(deviceId_).value;
   }
 
+  /**
+   * Get the forward-input value with deviceId.
+   */
+  const MatrixPtr& getInputValue(int inputIndex, int deviceId) {
+    return inputLayers_[inputIndex]->getOutput(deviceId).value;
+  }
+
   /**
    * Get the forward-input grad.
    */
@@ -200,6 +220,13 @@ protected:
     return inputLayer.getOutput(deviceId_).grad;
   }
 
+  /**
+   * Get the forward-input grad.
+   */
+  const MatrixPtr& getInputGrad(int inputIndex, int deviceId) {
+    return inputLayers_[inputIndex]->getOutput(deviceId).grad;
+  }
+
   /**
    * Get the forward-input label.
    */

paddle/gserver/layers/MKLDNNFcLayer.cpp

@@ -61,43 +61,42 @@ void MKLDNNFcLayer::convertWeightsFromPaddle() {
     return;
   }
 
-  // TODO(TJ): dst format should get from wgtVal_
-  int dstFmt = PARAM_FORMAT_MKLDNN_OI;
-  int srcFmt = weight_->getParameterPtr()->getHeaderFormat();
-  if (srcFmt == dstFmt) {
-    return;
-  }
-
-  // The weight_ is transposed from initial paddle weight
-  MatrixPtr paddleWgt = Matrix::create(
-      weight_->getW()->getData(), iLayerSize_, oc_, false, false);
-
-  // TODO(TJ): remove this print when do not need differ weights
-  std::ostringstream ostr;
-  paddleWgt->print(ostr);
-  VLOG(MKLDNN_ALL) << "Initial Weight from paddle: " << std::endl << ostr.str();
-
-  // The mkldnn weight is transposed from initial paddle matrix
-  MatrixPtr paddleWgtT;
-  paddleWgt->transpose(paddleWgtT, true);
-  weight_->getW()->copyFrom(*paddleWgtT);
-  weight_->getParameterPtr()->setHeaderFormat(dstFmt);
+  CHECK(wgtVal_) << "should have been initialized";
+  bool hasNoSpatial_ = ih_ == 1 && iw_ == 1;
+  auto targetDim = wgtVal_->getDims();
+  auto srcFmt = hasNoSpatial_ ? memory::format::io : memory::format::ihwo;
+  wgtVal_->reorderDataFrom(wgtVal_, srcFmt, targetDim);
   hasInitedWgt_ = true;
 }
 
 void MKLDNNFcLayer::convertWeightsToPaddle() {
-  MatrixPtr dnnWgt = weight_->getW();
-  MatrixPtr paddleWgt;
-  dnnWgt->transpose(paddleWgt, true);
-
-  // copy paddle weight and override on weight_
-  MatrixPtr dnnWgtT = Matrix::create(
-      dnnWgt->getData(), dnnWgt->getWidth(), dnnWgt->getHeight(), false, false);
-  dnnWgtT->copyFrom(*paddleWgt);
+  CHECK(wgtVal_) << "should have been initialized";
+  bool hasNoSpatial_ = ih_ == 1 && iw_ == 1;
+  auto targetDim = wgtVal_->getDims();
+  auto dstFmt = hasNoSpatial_ ? memory::format::io : memory::format::ihwo;
+  wgtVal_->reorderDataTo(wgtVal_, dstFmt, targetDim);
+}
+
+void MKLDNNFcLayer::convertOutputToOtherDevice() {
+  copyOutputInfoToOtherDevice();
+  // find other cpu device and reorder output to cpu device
+  int cnt = 0;
+  for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
+    if (outputOtherDevice_[i].deviceId == CPU_DEVICE) {
+      // fc cpu output value do not need convert
+      // just share point
+      outputOtherDevice_[i].value = output_.value;
+      ++cnt;
+    }
+  }
+
+  if (cnt > 1) {
+    LOG(WARNING) << "should not have more than one CPU devie";
+  }
 }
 
 void MKLDNNFcLayer::reshape() {
-  const Argument& input = getInput(0);
+  const Argument& input = getInput(0, getPrev(0)->getDeviceId());
   int batchSize = input.getBatchSize();
   if (bs_ == batchSize) {
     return;
@@ -111,10 +110,6 @@ void MKLDNNFcLayer::reshape() {
   if (iw_ == 0) {
     iw_ = 1;
   }
-  hasSpatial_ = true;
-  if (ih_ == 1 && iw_ == 1) {
-    hasSpatial_ = false;
-  }
   CHECK_EQ(iLayerSize_, inputLayers_[0]->getSize());
   ic_ = iLayerSize_ / (ih_ * iw_);
   CHECK_EQ(size_t(ic_ * ih_ * iw_), iLayerSize_) << "not divisible";
@@ -135,37 +130,53 @@ void MKLDNNFcLayer::reshape() {
 
 void MKLDNNFcLayer::resetFwd() {
   bool hasBias = biases_ && biases_->getW();
-  real* iData = getInputValue(0)->getData();
-  real* oData = getOutputValue()->getData();
-  real* wData = weight_->getW()->getData();
-  real* bData = hasBias ? biases_->getW()->getData() : NULL;
-
-  // TODO(TJ): below create should be covered in MkldnnMatrix
-  // create memory desc
-  memory::desc iMD = hasSpatial_ ? createMD({bs_, ic_, ih_, iw_}, format::nchw)
-                                 : createMD({bs_, ic_}, format::nc);
-  memory::desc wMD = hasSpatial_ ? createMD({oc_, ic_, ih_, iw_}, format::oihw)
-                                 : createMD({oc_, ic_}, format::oi);
-  memory::desc bMD = bData != NULL ? createMD({oc_}, format::x)
-                                   : createMD({}, format::format_undef);
-  memory::desc oMD = createMD({bs_, oc_}, format::nc);
-
-  // create memory primitive desc and memory self
-  inVal_.reset(new memory(memory::primitive_desc(iMD, engine_), iData));
-  wgtVal_.reset(new memory(memory::primitive_desc(wMD, engine_), wData));
-  outVal_.reset(new memory(memory::primitive_desc(oMD, engine_), oData));
+  const MatrixPtr& wgt = weight_->getW();
+  const MatrixPtr& bias = hasBias ? biases_->getW() : nullptr;
+  const MatrixPtr& out = output_.value;
+
+  if (inputIsOnlyMKLDNN()) {
+    const MatrixPtr& in = getInputValue(0);
+    inVal_ = std::dynamic_pointer_cast<MKLDNNMatrix>(in);
+    CHECK(inVal_) << "Input should be MKLDNNMatrix";
+  } else {
+    CHECK_EQ(getPrev(0)->getDeviceId(), CPU_DEVICE) << "Only support CPU yet";
+    const MatrixPtr& in = getInputValue(0, CPU_DEVICE);
+    inVal_ = MKLDNNMatrix::create(
+        in, memory::dims{bs_, ic_, ih_, iw_}, format::nchw, engine_);
+  }
+  inVal_->downSpatial();
+  wgtVal_ = MKLDNNMatrix::create(
+      wgt, memory::dims{oc_, ic_, ih_, iw_}, format::oihw, engine_);
+  wgtVal_->downSpatial();
+  biasVal_ =
+      hasBias ? MKLDNNMatrix::create(bias, {oc_}, format::x, engine_) : nullptr;
+  outVal_ = MKLDNNMatrix::create(out, {bs_, oc_}, format::nc, engine_);
+
+  // change original output value to mkldnn output value
+  output_.value = std::dynamic_pointer_cast<Matrix>(outVal_);
+  if (!outputIsOnlyMKLDNN()) {
+    convertOutputToOtherDevice();
+  }
 
+  // create forward handle
   prop_kind pk = prop_kind::forward;
-  fc_fwd::desc fwdDesc = bData != NULL ? fc_fwd::desc(pk, iMD, wMD, bMD, oMD)
-                                       : fc_fwd::desc(pk, iMD, wMD, oMD);
+  fc_fwd::desc fwdDesc = hasBias ? fc_fwd::desc(pk,
+                                                inVal_->getMemoryDesc(),
+                                                wgtVal_->getMemoryDesc(),
+                                                biasVal_->getMemoryDesc(),
+                                                outVal_->getMemoryDesc())
+                                 : fc_fwd::desc(pk,
+                                                inVal_->getMemoryDesc(),
+                                                wgtVal_->getMemoryDesc(),
+                                                outVal_->getMemoryDesc());
   fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_);
-
-  if (bData != NULL) {
-    biasVal_.reset(new memory(memory::primitive_desc(bMD, engine_), bData));
+  if (hasBias) {
     fwd_.reset(new fc_fwd(fwdPD, *inVal_, *wgtVal_, *biasVal_, *outVal_));
   } else {
     fwd_.reset(new fc_fwd(fwdPD, *inVal_, *wgtVal_, *outVal_));
   }
+  printValueFormatFlow();
+
   pipelineFwd_.clear();
   pipelineFwd_.push_back(*fwd_);
 }
@@ -175,45 +186,46 @@ void MKLDNNFcLayer::resetBwd() {
     return;
   }
   needResetBwd_ = false;
-
   bool hasBias = biases_ && biases_->getWGrad();
-  real* iData = getInputValue(0)->getData();
-  real* iDiff = getInputGrad(0) != nullptr ? getInputGrad(0)->getData() : NULL;
-  real* oDiff = getOutputGrad()->getData();
-  real* wDiff = weight_->getWGrad()->getData();
-  real* bDiff = hasBias ? biases_->getWGrad()->getData() : NULL;
 
   /// backward weight
-  // create memory desc for backward memory
-  memory::desc iMD = hasSpatial_ ? createMD({bs_, ic_, ih_, iw_}, format::nchw)
-                                 : createMD({bs_, ic_}, format::nc);
-  memory::desc wMD = hasSpatial_ ? createMD({oc_, ic_, ih_, iw_}, format::oihw)
-                                 : createMD({oc_, ic_}, format::oi);
-  memory::desc oMD = createMD({bs_, oc_}, format::nc);
-  memory::desc bMD = bDiff != NULL ? createMD({oc_}, format::x)
-                                   : createMD({}, format::format_undef);
-
-  if (inVal_) {
-    // update data
-    inVal_->set_data_handle(iData);
-  } else {
-    inVal_.reset(new memory(memory::primitive_desc(iMD, engine_), iData));
-  }
-
-  // create memory primitive desc and memory self
-  wgtGrad_.reset(new memory(memory::primitive_desc(wMD, engine_), wDiff));
-  outGrad_.reset(new memory(memory::primitive_desc(oMD, engine_), oDiff));
-
-  fc_fwd::desc fwdDesc = fc_fwd::desc(prop_kind::forward, iMD, wMD, oMD);
+  CHECK(inVal_) << "Should have input value";
+  const MatrixPtr& wgt = weight_->getWGrad();
+  const MatrixPtr& bias = hasBias ? biases_->getWGrad() : nullptr;
+
+  // TODO(TJ): merge outgrad
+  int device = outputIsOnlyMKLDNN() ? MKLDNN_DEVICE : CPU_DEVICE;
+  // for MKLDNN device:
+  // can not directly cast outputgrad to mkldnnmatrix,
+  // since each layer can not write the inputgrad to mkldnn inputgrad.
+  // So just create from matrix with outputvalue format.
+  // for CPU device:
+  // fc do not need to convert from cpu device since output is always nc format
+  // only need create from cpu device
+  const MatrixPtr& out = getOutput(device).grad;
+  outGrad_ = MKLDNNMatrix::create(out, outVal_->getPrimitiveDesc());
+  wgtGrad_ = MKLDNNMatrix::create(wgt, wgtVal_->getPrimitiveDesc());
+  biasGrad_ = hasBias ? MKLDNNMatrix::create(bias, biasVal_->getPrimitiveDesc())
+                      : nullptr;
+
+  // create memory primitive desc
+  fc_fwd::desc fwdDesc = fc_fwd::desc(prop_kind::forward,
+                                      inVal_->getMemoryDesc(),
+                                      wgtGrad_->getMemoryDesc(),
+                                      outGrad_->getMemoryDesc());
   fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_);
-  fc_bwdWgt::desc bwdWgtDesc = bDiff != NULL
-                                   ? fc_bwdWgt::desc(iMD, wMD, bMD, oMD)
-                                   : fc_bwdWgt::desc(iMD, wMD, oMD);
+  fc_bwdWgt::desc bwdWgtDesc = hasBias
+                                   ? fc_bwdWgt::desc(inVal_->getMemoryDesc(),
+                                                     wgtGrad_->getMemoryDesc(),
+                                                     biasGrad_->getMemoryDesc(),
+                                                     outGrad_->getMemoryDesc())
+                                   : fc_bwdWgt::desc(inVal_->getMemoryDesc(),
+                                                     wgtGrad_->getMemoryDesc(),
+                                                     outGrad_->getMemoryDesc());
   fc_bwdWgt::primitive_desc bwdWgtPD =
       fc_bwdWgt::primitive_desc(bwdWgtDesc, engine_, fwdPD);
 
-  if (bDiff != NULL) {
-    biasGrad_.reset(new memory(memory::primitive_desc(bMD, engine_), bDiff));
+  if (hasBias) {
     bwdWgt_.reset(
         new fc_bwdWgt(bwdWgtPD, *inVal_, *outGrad_, *wgtGrad_, *biasGrad_));
   } else {
@@ -223,15 +235,26 @@ void MKLDNNFcLayer::resetBwd() {
   pipelineBwd_.push_back(*bwdWgt_);
 
   /// backward data
-  if (iDiff == NULL) {
+  device = inputIsOnlyMKLDNN() ? MKLDNN_DEVICE : CPU_DEVICE;
+  const MatrixPtr& in = getInputGrad(0, device);
+  if (in == nullptr) {
     return;
   }
-  fc_bwdData::desc bwdDataDesc = fc_bwdData::desc(iMD, wMD, oMD);
+  if (getInput(0, device).getAllCount() > 1) {
+    // TODO(TJ): use outputMaps_ ways when merge outgrad done
+  } else {
+    inGrad_ = MKLDNNMatrix::create(in, inVal_->getPrimitiveDesc());
+  }
+
+  fc_bwdData::desc bwdDataDesc = fc_bwdData::desc(inVal_->getMemoryDesc(),
+                                                  wgtGrad_->getMemoryDesc(),
+                                                  outGrad_->getMemoryDesc());
   fc_bwdData::primitive_desc bwdDataPD =
       fc_bwdData::primitive_desc(bwdDataDesc, engine_, fwdPD);
-  inGrad_.reset(new memory(memory::primitive_desc(iMD, engine_), iDiff));
+
   CHECK(wgtVal_) << "Should have weight memory";
   bwdData_.reset(new fc_bwdData(bwdDataPD, *outGrad_, *wgtVal_, *inGrad_));
+  printGradFormatFlow();
   pipelineBwd_.push_back(*bwdData_);
 }
 
@@ -241,11 +264,7 @@ void MKLDNNFcLayer::forward(PassType passType) {
 
   {
     REGISTER_TIMER_INFO("mkldnn_FwdTimer", getName().c_str());
-
-    // update input data
-    // since it might be changed if this is after data layer
-    real* iData = getInputValue(0)->getData();
-    inVal_->set_data_handle(iData);
+    syncInputValue();
 
     // just submit forward pipeline
     stream_->submit(pipelineFwd_);
@@ -267,10 +286,7 @@ void MKLDNNFcLayer::backward(const UpdateCallback& callback) {
     REGISTER_TIMER_INFO("mkldnn_bwdTimer", getName().c_str());
     resetBwd();
 
-    // update diff
-    real* oDiff = getOutputGrad()->getData();
-    outGrad_->set_data_handle(oDiff);
-
+    syncOutputGrad();
     // just sumbmit backward pipeline
     stream_->submit(pipelineBwd_);
   }

paddle/gserver/layers/MKLDNNFcLayer.h

@@ -32,16 +32,13 @@ protected:
   // if has already init the weight
   bool hasInitedWgt_;
 
-  // if input layer has image size info (ih>1 && iw>1)
-  bool hasSpatial_;
-
   // fc weight and bias
   std::unique_ptr<Weight> weight_;
   std::unique_ptr<Weight> biases_;
 
 public:
   explicit MKLDNNFcLayer(const LayerConfig& config)
-      : MKLDNNLayer(config), hasInitedWgt_(false), hasSpatial_(true) {}
+      : MKLDNNLayer(config), hasInitedWgt_(false) {}
 
   ~MKLDNNFcLayer() {}
 
@@ -75,6 +72,8 @@ protected:
    * only would be called when needed
    */
   void resetBwd();
+
+  void convertOutputToOtherDevice() override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/MKLDNNLayer.h

@@ -18,9 +18,9 @@ limitations under the License. */
 #include "Layer.h"
 #include "MKLDNNBase.h"
 #include "mkldnn.hpp"
+#include "paddle/math/MKLDNNMatrix.h"
 
 DECLARE_bool(use_mkldnn);
-DECLARE_bool(use_mkldnn_wgt);
 
 namespace paddle {
 
@@ -52,15 +52,15 @@ protected:
   std::vector<mkldnn::primitive> pipelineFwd_;
   std::vector<mkldnn::primitive> pipelineBwd_;
 
-  // TODO(TJ): change below memory as MKLDNNMatrixPtr type
-  std::shared_ptr<mkldnn::memory> inVal_;
-  std::shared_ptr<mkldnn::memory> inGrad_;
-  std::shared_ptr<mkldnn::memory> outVal_;
-  std::shared_ptr<mkldnn::memory> outGrad_;
-  std::shared_ptr<mkldnn::memory> wgtVal_;
-  std::shared_ptr<mkldnn::memory> wgtGrad_;
-  std::shared_ptr<mkldnn::memory> biasVal_;
-  std::shared_ptr<mkldnn::memory> biasGrad_;
+  // MKLDNNMatrixPtr
+  MKLDNNMatrixPtr inVal_;
+  MKLDNNMatrixPtr inGrad_;
+  MKLDNNMatrixPtr outVal_;
+  MKLDNNMatrixPtr outGrad_;
+  MKLDNNMatrixPtr wgtVal_;
+  MKLDNNMatrixPtr wgtGrad_;
+  MKLDNNMatrixPtr biasVal_;
+  MKLDNNMatrixPtr biasGrad_;
 
 public:
   explicit MKLDNNLayer(const LayerConfig& config)
@@ -83,17 +83,21 @@ public:
 
   virtual bool init(const LayerMap& layerMap,
                     const ParameterMap& parameterMap) {
+    CHECK(FLAGS_use_mkldnn) << "MkldnnLayers only support use_mkldnn."
+                            << "Please set WITH_MKLDNN=ON "
+                            << "and set use_mkldnn=True";
+    CHECK(!useGpu_) << "Do not support GPU yet";
+
+    // set device id before Layer::init
+    setDevice(MKLDNN_DEVICE);
+    // change param device to MKLDNN device
+    setParamsDevice(MKLDNN_DEVICE, parameterMap);
     if (!Layer::init(layerMap, parameterMap)) {
       return false;
     }
 
-    CHECK(FLAGS_use_mkldnn) << "MkldnnLayers only support use_mkldnn."
-                            << "Please set WITH_MKLDNN=ON "
-                            << "and set use_mkldnn=True";
     stream_.reset(new MKLDNNStream());
     engine_ = CPUEngine::Instance().getEngine();
-
-    // TODO(TJ): deivecId
     return true;
   }
 
@@ -109,6 +113,12 @@ public:
    */
   virtual void convertWeightsToPaddle() {}
 
+  /**
+   * convert MKLDNN output to other device.
+   * only support CPU device yet
+   */
+  virtual void convertOutputToOtherDevice() {}
+
   /**
    * print info about sizes
    */
@@ -118,14 +128,124 @@ public:
                        << ", oh: " << oh_ << ", ow: " << ow_;
   }
 
-  // TODO(TJ): move to MkldnnMatrix
-  // create memory desc
-  inline mkldnn::memory::desc createMD(
-      mkldnn::memory::dims dims,
-      mkldnn::memory::format fmt,
-      mkldnn::memory::data_type type = mkldnn::memory::data_type::f32) {
-    // TODO(TJ): isFmtSuppoted(fmt)
-    return mkldnn::memory::desc(dims, type, fmt);
+  /**
+   * Print the mkldnn memory format flow of value
+   */
+  virtual void printValueFormatFlow() {
+    if (inVal_ && outVal_) {
+      VLOG(MKLDNN_FMTS) << "value format flow --- " << inVal_->getFormat()
+                        << " >>> " << outVal_->getFormat();
+    }
+  }
+
+  /**
+   * Print the mkldnn memory format flow of grad
+   */
+  virtual void printGradFormatFlow() {
+    if (inGrad_ && outGrad_) {
+      VLOG(MKLDNN_FMTS) << "grad format flow --- " << inGrad_->getFormat()
+                        << " <<< " << outGrad_->getFormat();
+    }
+  }
+
+protected:
+  /**
+   * copy image size and sequence info to other device
+   * @note: can not directly use Layer::copyOutputToOtherDevice since here only
+   *        copy base info and do not copy data value
+   */
+  void copyOutputInfoToOtherDevice() {
+    for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
+      outputOtherDevice_[i].setFrameHeight(output_.getFrameHeight());
+      outputOtherDevice_[i].setFrameWidth(output_.getFrameWidth());
+      outputOtherDevice_[i].sequenceStartPositions =
+          output_.sequenceStartPositions;
+      outputOtherDevice_[i].subSequenceStartPositions =
+          output_.subSequenceStartPositions;
+      outputOtherDevice_[i].cpuSequenceDims = output_.cpuSequenceDims;
+    }
+  }
+
+  /**
+   * If input only has MKLDNN device.
+   * Otherwise, only support the previous layer using CPU device.
+   */
+  bool inputIsOnlyMKLDNN(int index = 0) {
+    int prevDevice = getPrev(index)->getDeviceId();
+    if (prevDevice == MKLDNN_DEVICE) {
+      return true;
+    } else {
+      // do not support GPU yet
+      CHECK_EQ(prevDevice, CPU_DEVICE) << "Only support CPU yet";
+      return false;
+    }
+  }
+
+  /**
+   * If output only has MKLDNN device.
+   * Otherwise, other devices should only using CPU device.
+   */
+  bool outputIsOnlyMKLDNN() {
+    for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
+      CHECK_EQ(outputOtherDevice_[i].deviceId, CPU_DEVICE)
+          << "Only support other device is CPU yet";
+    }
+    return outputOtherDevice_.size() == 0;
+  }
+
+  /**
+   * Sync input value data
+   */
+  void syncInputValue() {
+    if (inputIsOnlyMKLDNN()) {
+      return;
+    }
+    real* iData = getInputValue(0, CPU_DEVICE)->getData();
+    // update input data
+    // since it might be changed if this is after data layer
+    inVal_->updateData(iData);
+  }
+
+  /**
+   * Sync output grad data
+   */
+  void syncOutputGrad() {
+    if (outputIsOnlyMKLDNN()) {
+      return;
+    }
+
+    // update diff
+    real* oDiff = getOutput(CPU_DEVICE).grad->getData();
+    outGrad_->updateData(oDiff);
+  }
+
+  /**
+   * Set deviceId of this layer.
+   */
+  void setDevice(int id) { deviceId_ = id; }
+
+  /**
+   * Set deviceId of the params used in this layer.
+   */
+  void setParamsDevice(int id, const ParameterMap& parameterMap) {
+    for (auto& inputConfig : config_.inputs()) {
+      if (inputConfig.has_input_parameter_name()) {
+        ParameterPtr parameter;
+        std::string name = inputConfig.input_parameter_name();
+        CHECK(mapGet(name, parameterMap, &parameter))
+            << "Cannot find input parameter " << name << " for layer "
+            << getName();
+        parameter->setDevice(id);
+      }
+    }
+    if (config_.has_bias_parameter_name()) {
+      ParameterPtr parameter;
+      std::string name = config_.bias_parameter_name();
+      CHECK(mapGet(name, parameterMap, &parameter))
+          << "Cannot find bias parameter " << name << " for layer "
+          << getName();
+      parameter->setDevice(id);
+    }
   }
 };
 

paddle/gserver/layers/Pool3DLayer.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 "Pool3DLayer.h"
+#include "PoolProjectionLayer.h"
+#include "paddle/utils/Logging.h"
+
+namespace paddle {
+
+REGISTER_LAYER(pool3d, Pool3DLayer);
+
+bool Pool3DLayer::init(const LayerMap& layerMap,
+                       const ParameterMap& parameterMap) {
+  Layer::init(layerMap, parameterMap);
+
+  /* the size of inputs for pool-layer is 1 */
+  CHECK_EQ(config_.inputs_size(), 1);
+
+  const PoolConfig& conf = config_.inputs(0).pool_conf();
+  poolType_ = conf.pool_type();
+  channels_ = conf.channels();
+
+  sizeX_ = conf.size_x();
+  sizeY_ = conf.size_y();
+  sizeZ_ = conf.size_z();
+
+  strideW_ = conf.stride();
+  strideH_ = conf.stride_y();
+  strideD_ = conf.stride_z();
+
+  imgSizeW_ = conf.img_size();
+  imgSizeH_ = conf.img_size_y();
+  imgSizeD_ = conf.img_size_z();
+
+  paddingW_ = conf.padding();
+  paddingH_ = conf.padding_y();
+  paddingD_ = conf.padding_z();
+
+  outputW_ = conf.output_x();
+  outputH_ = conf.output_y();
+  outputD_ = conf.output_z();
+
+  return true;
+}
+
+size_t Pool3DLayer::getSize() {
+  CHECK_EQ(inputLayers_.size(), 1UL);
+
+  size_t layerSize = 0;
+  outputD_ = outputSize(imgSizeD_, sizeZ_, paddingD_, strideD_, false);
+  outputH_ = outputSize(imgSizeH_, sizeY_, paddingH_, strideH_, false);
+  outputW_ = outputSize(imgSizeW_, sizeX_, paddingW_, strideW_, false);
+
+  layerSize = outputD_ * outputH_ * outputW_ * channels_;
+  getOutput().setFrameHeight(outputH_);
+  getOutput().setFrameWidth(outputW_);
+  getOutput().setFrameDepth(outputD_);
+  return layerSize;
+}
+
+void Pool3DLayer::forward(PassType passType) {
+  Layer::forward(passType);
+  const MatrixPtr& inMat = inputLayers_[0]->getOutputValue();
+  size_t batchSize = inMat->getHeight();
+  size_t outWidth = getSize();
+  resetOutput(batchSize, outWidth);
+  Matrix::resizeOrCreate(maxPoolIdx_, batchSize, outWidth, false, useGpu_);
+  const MatrixPtr outMat = getOutputValue();
+
+  if (poolType_ == "avg") {
+    outMat->avgPool3DForward(*inMat,
+                             channels_,
+                             imgSizeD_,
+                             imgSizeH_,
+                             imgSizeW_,
+                             outputD_,
+                             outputH_,
+                             outputW_,
+                             sizeZ_,
+                             sizeY_,
+                             sizeX_,
+                             strideD_,
+                             strideH_,
+                             strideW_,
+                             paddingD_,
+                             paddingH_,
+                             paddingW_);
+  } else if (poolType_ == "max") {
+    outMat->maxPool3DForward(*inMat,
+                             *maxPoolIdx_,
+                             channels_,
+                             imgSizeD_,
+                             imgSizeH_,
+                             imgSizeW_,
+                             outputD_,
+                             outputH_,
+                             outputW_,
+                             sizeZ_,
+                             sizeY_,
+                             sizeX_,
+                             strideD_,
+                             strideH_,
+                             strideW_,
+                             paddingD_,
+                             paddingH_,
+                             paddingW_);
+  } else {
+    LOG(FATAL) << "Unknown pool type: " << poolType_;
+  }
+  forwardActivation();
+}
+
+void Pool3DLayer::backward(const UpdateCallback& callback) {
+  backwardActivation();
+
+  (void)callback;
+  if (NULL == getInputGrad(0)) return;
+  MatrixPtr inMat = inputLayers_[0]->getOutputValue();
+  MatrixPtr inGradMat = inputLayers_[0]->getOutputGrad();
+  MatrixPtr outMat = getOutputValue();
+  MatrixPtr outGradMat = getOutputGrad();
+
+  if (poolType_ == "avg") {
+    inGradMat->avgPool3DBackward(*outGradMat,
+                                 imgSizeD_,
+                                 imgSizeH_,
+                                 imgSizeW_,
+                                 outputD_,
+                                 outputH_,
+                                 outputW_,
+                                 sizeZ_,
+                                 sizeY_,
+                                 sizeZ_,
+                                 strideD_,
+                                 strideH_,
+                                 strideW_,
+                                 paddingD_,
+                                 paddingH_,
+                                 paddingW_,
+                                 1.0,
+                                 1.0);
+  } else if (poolType_ == "max") {
+    inGradMat->maxPool3DBackward(*outGradMat,
+                                 *maxPoolIdx_,
+                                 imgSizeD_,
+                                 imgSizeH_,
+                                 imgSizeW_,
+                                 outputD_,
+                                 outputH_,
+                                 outputW_,
+                                 sizeZ_,
+                                 sizeY_,
+                                 sizeZ_,
+                                 strideD_,
+                                 strideH_,
+                                 strideW_,
+                                 paddingD_,
+                                 paddingH_,
+                                 paddingW_,
+                                 1.0,
+                                 1.0);
+  } else {
+    LOG(FATAL) << "Unknown pool type: " << poolType_;
+  }
+}
+
+}  // namespace paddle

paddle/gserver/layers/Pool3DLayer.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 "Layer.h"
+#include "paddle/math/MathUtils.h"
+#include "paddle/math/Matrix.h"
+
+namespace paddle {
+
+/**
+ * @brief Basic parent layer of pooling
+ * Pools the input within regions
+ */
+class Pool3DLayer : public Layer {
+public:
+  explicit Pool3DLayer(const LayerConfig& config) : Layer(config) {}
+  ~Pool3DLayer() {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback) override;
+  size_t getSize();
+
+protected:
+  int channels_;
+  int sizeX_, sizeY_, sizeZ_;
+  int strideW_, strideH_, strideD_;
+  int paddingW_, paddingH_, paddingD_;
+  int imgSizeW_, imgSizeH_, imgSizeD_;
+  int outputW_, outputH_, outputD_;
+  std::string poolType_;
+  MatrixPtr maxPoolIdx_;
+};
+}  // namespace paddle

paddle/gserver/layers/PrintLayer.cpp

@@ -48,7 +48,16 @@ public:
                  << inputLayers_.size() << ") at " << getName();
     }
     s << format.substr(pos);
-    LOG(INFO) << s.str();
+
+    const std::string delimiter("\n");
+    std::string content = s.str();
+    std::string::size_type foundPos = 0;
+    std::string::size_type prevPos = 0;
+    while ((foundPos = content.find(delimiter, prevPos)) != std::string::npos) {
+      LOG(INFO) << content.substr(prevPos, foundPos - prevPos);
+      prevPos = foundPos + delimiter.size();
+    }
+    LOG(INFO) << content.substr(prevPos);
   }
 
   void backward(const UpdateCallback& callback) override {}

paddle/gserver/tests/CMakeLists.txt

@@ -34,6 +34,13 @@ add_unittest_without_exec(test_CRFLayerGrad
 add_test(NAME test_CRFLayerGrad
     COMMAND test_CRFLayerGrad)
 
+################ test_CrossEntropyOverBeam ####################
+add_unittest_without_exec(test_CrossEntropyOverBeam
+    test_CrossEntropyOverBeamGrad.cpp
+    LayerGradUtil.cpp)
+add_test(NAME test_CrossEntropyOverBeam
+    COMMAND test_CrossEntropyOverBeam)
+
 ################ test_SeqSliceLayerGrad ####################
 add_unittest_without_exec(test_SeqSliceLayerGrad
     test_SeqSliceLayerGrad.cpp

paddle/gserver/tests/test_CrossEntropyOverBeamGrad.cpp

+/* Copyright (c) 2016 Baidu, Inc. 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 <random>
+#include <sstream>
+
+#include <gtest/gtest.h>
+#include "ModelConfig.pb.h"
+#include "paddle/gserver/layers/DataLayer.h"
+#include "paddle/trainer/Trainer.h"
+
+#include "LayerGradUtil.h"
+#include "paddle/testing/TestUtil.h"
+
+using namespace paddle;  // NOLINT
+
+DECLARE_int32(gpu_id);
+DECLARE_bool(thread_local_rand_use_global_seed);
+
+const size_t MAX_SEQ_NUM = 23;
+const size_t MAX_SEQ_LEN = 50;
+const size_t MAX_BEAM_SIZE = 27;
+
+const size_t SEED = (size_t)(time(NULL));
+
+struct SingleBeamExpansion {
+  vector<int> seqStartPos;
+  vector<int> subSeqStartPos;
+  vector<real> candidateScores;
+
+  // TODO(caoying): store this into Argument.ids
+  vector<real> selectedIndices;
+
+  vector<int> groundTruth;
+  vector<size_t> inBeam;
+  vector<int> rowIdxInBeam;
+  vector<int> colIdxInBeam;
+
+  void resetGroundTruth(size_t n) {
+    groundTruth.clear();
+    groundTruth.resize(n, -1);
+
+    inBeam.clear();
+    inBeam.resize(n, 0);
+
+    rowIdxInBeam.clear();
+    rowIdxInBeam.resize(n, -1);
+
+    colIdxInBeam.clear();
+    colIdxInBeam.resize(n, -1);
+  }
+};
+
+inline float randFloat() {
+  return static_cast<float>(rand()) / static_cast<float>(RAND_MAX);
+}
+
+void genRand(real* numbers, size_t n) {
+  default_random_engine generator;
+  uniform_real_distribution<real> distribution(0.0, 1.0);
+  for (size_t i = 0; i < n; ++i) numbers[i] = distribution(generator);
+}
+
+vector<real> randSampling(real range, int n) {
+  CHECK_GE(range, n);
+  vector<real> num(range);
+  iota(begin(num), end(num), 0.);
+  if (range == n) return num;
+
+  random_shuffle(begin(num), end(num));
+  num.resize(n);
+  sort(begin(num), end(num));
+  return num;
+}
+
+void genCandidateScores(bool hasSubseq,
+                        size_t beamSize,
+                        SingleBeamExpansion& prevBeam,
+                        SingleBeamExpansion& curBeam) {
+  vector<int>& seqStartPos = curBeam.seqStartPos;
+  seqStartPos.resize(1, 0);
+  vector<int>& subSeqStartPos = curBeam.subSeqStartPos;
+  subSeqStartPos.resize(1, 0);
+
+  srand(SEED);
+  if (prevBeam.selectedIndices.size()) {
+    if (prevBeam.subSeqStartPos.size() > 1) {
+      int seqIdx = 1;
+      // samples in previous beam are nested sequences.
+      for (size_t i = 1; i < prevBeam.subSeqStartPos.size(); ++i) {
+        for (size_t j = 0; j < beamSize; ++j) {
+          if (prevBeam.selectedIndices[(i - 1) * beamSize + j] == -1.) break;
+          subSeqStartPos.push_back(1 + (rand() % MAX_SEQ_LEN) +
+                                   subSeqStartPos.back());
+        }
+        if (prevBeam.seqStartPos[seqIdx] == prevBeam.subSeqStartPos[i]) {
+          seqStartPos.push_back(subSeqStartPos.back());
+          seqIdx++;
+        }
+      }
+    } else {
+      for (size_t i = 0; i <= prevBeam.selectedIndices.size(); ++i) {
+        if (i && i % beamSize == 0) {
+          seqStartPos.push_back(subSeqStartPos.back());
+          if (i == prevBeam.selectedIndices.size()) break;
+        }
+        if (prevBeam.selectedIndices[i] == -1.) continue;
+        subSeqStartPos.push_back(subSeqStartPos.back() +
+                                 (1 + (rand() % MAX_SEQ_LEN)));
+      }
+    }
+  } else {
+    // the first beam expansion
+    int seqNum = 1 + (rand() % MAX_SEQ_NUM);
+    for (int i = 0; i < seqNum; ++i) {
+      if (hasSubseq) {
+        for (size_t j = 0; j < 1 + (rand() % MAX_SEQ_NUM); ++j)
+          subSeqStartPos.push_back(subSeqStartPos.back() +
+                                   (1 + (rand() % MAX_SEQ_LEN)));
+        seqStartPos.push_back(subSeqStartPos.back());
+      } else {
+        seqStartPos.push_back(seqStartPos.back() +
+                              (1 + (rand() % MAX_SEQ_LEN)));
+      }
+    }
+  }
+
+  size_t totalSeqNum = hasSubseq ? subSeqStartPos.back() : seqStartPos.back();
+  curBeam.candidateScores.resize(totalSeqNum, 0.);
+  genRand(curBeam.candidateScores.data(), totalSeqNum);
+}
+
+void genSelectedIndices(size_t beamSize,
+                        vector<int>& seqStartPos,
+                        vector<real>& selectedIndices) {
+  size_t selectedIdsCount = beamSize * (seqStartPos.size() - 1);
+  selectedIndices.resize(selectedIdsCount, -1.);
+
+  for (size_t i = 0; i < seqStartPos.size() - 1; ++i) {
+    int seqLen = seqStartPos[i + 1] - seqStartPos[i];
+    int n = min(seqLen, static_cast<int>(beamSize));
+    vector<real> ids = randSampling(seqLen, n);
+    memcpy(selectedIndices.data() + i * beamSize,
+           ids.data(),
+           sizeof(real) * ids.size());
+  }
+}
+
+void genGroundTruth(vector<SingleBeamExpansion>& beamExpansions,
+                    size_t beamSize) {
+  SingleBeamExpansion& beam = beamExpansions[1];
+  size_t seqNum = beam.seqStartPos.size() - 1;
+  for (size_t i = 2; i < beamExpansions.size(); ++i)
+    CHECK_EQ(seqNum, beamExpansions[i].seqStartPos.size() - 1);
+
+  srand(SEED);
+
+  // initialize the first beam.
+  beam.resetGroundTruth(seqNum);
+  for (size_t i = 0; i < seqNum; ++i) {
+    if (randFloat() > 0.5) {
+      /*
+       * force the randomly generated label falls in the beam by chance 0.5.
+       * otherwise, when sequence length is relatively long and beam size is
+       * relatively small, the gold sequences falls off the beam at in the
+       * first search.
+       */
+      real* begPos = beam.selectedIndices.data() + i * beamSize;
+      beam.colIdxInBeam[i] =
+          rand() % count_if(begPos, begPos + beamSize, [](const real& val) {
+            return val != -1.;
+          });
+      beam.groundTruth[i] =
+          beam.selectedIndices[i * beamSize + beam.colIdxInBeam[i]];
+      beam.inBeam[i] = 1;
+    } else {
+      int label = rand() % (beam.seqStartPos[i + 1] - beam.seqStartPos[i]);
+      beam.groundTruth[i] = label;
+
+      real* begPos = beam.selectedIndices.data() + i * beamSize;
+      real* endPos = begPos + beamSize;
+      real* lblPos = find(begPos, endPos, real(label));
+      if (lblPos != endPos) {
+        beam.inBeam[i] = 1;
+        beam.colIdxInBeam[i] = lblPos - begPos;
+      }
+    }
+    beam.rowIdxInBeam[i] = i;
+  }
+
+  // iterate over each beam expansions
+  for (size_t i = 2; i < beamExpansions.size(); ++i) {
+    SingleBeamExpansion& curBeam = beamExpansions[i];
+    SingleBeamExpansion& prevBeam = beamExpansions[i - 1];
+    curBeam.resetGroundTruth(seqNum);
+
+    // iterate over each sequence
+    for (size_t j = 0; j < seqNum; ++j) {
+      if (!prevBeam.inBeam[j]) continue;
+
+      // gold sequence falls in the beam in previous search.
+      real* begPos = prevBeam.selectedIndices.data();
+      int offset =
+          prevBeam.rowIdxInBeam[j] * beamSize + prevBeam.colIdxInBeam[j];
+      curBeam.rowIdxInBeam[j] = count_if(
+          begPos, begPos + offset, [](const real& val) { return val != -1.; });
+
+      if (randFloat() > 0.5) {
+        // force the randomly generated label falls in the beam by chance 0.5.
+
+        real* start =
+            curBeam.selectedIndices.data() + curBeam.rowIdxInBeam[j] * beamSize;
+        int n = rand() % count_if(start, start + beamSize, [](const real& val) {
+                  return val != -1.;
+                });
+        curBeam.colIdxInBeam[j] = n;
+        curBeam.groundTruth[j] = *(start + n);
+        curBeam.inBeam[j] = 1;
+      } else {
+        CHECK_LE(curBeam.rowIdxInBeam[j] + 1,
+                 curBeam.subSeqStartPos.size() - 1);
+        int start = curBeam.subSeqStartPos[curBeam.rowIdxInBeam[j]];
+        int end = curBeam.subSeqStartPos[curBeam.rowIdxInBeam[j] + 1];
+        CHECK_GT(size_t(end), size_t(start));
+        int label = rand() % (end - start);
+
+        curBeam.groundTruth[j] = label;
+        real* findBeg =
+            curBeam.selectedIndices.data() + curBeam.rowIdxInBeam[j] * beamSize;
+        real* lblPos =
+            find(findBeg, findBeg + beamSize, static_cast<real>(label));
+        if (lblPos != (findBeg + beamSize)) {
+          curBeam.inBeam[j] = 1;
+          curBeam.colIdxInBeam[j] = lblPos - findBeg;
+        }
+      }
+    }
+  }
+}
+
+void genOneBeam(size_t beamSize,
+                bool hasSubseq,
+                SingleBeamExpansion& prevBeam,
+                SingleBeamExpansion& curBeam) {
+  genCandidateScores(hasSubseq, beamSize, prevBeam, curBeam);
+  genSelectedIndices(beamSize,
+                     hasSubseq ? curBeam.subSeqStartPos : curBeam.seqStartPos,
+                     curBeam.selectedIndices);
+}
+
+void genRandomBeamExpansion(size_t expansionCount,
+                            size_t beamSize,
+                            vector<SingleBeamExpansion>& beamExpansions) {
+  beamExpansions.clear();
+  beamExpansions.resize(expansionCount + 1);
+
+  // beamExpansions[0] is reserved.
+  for (size_t i = 1; i <= expansionCount; ++i)
+    genOneBeam(beamSize, bool(i - 1), beamExpansions[i - 1], beamExpansions[i]);
+  genGroundTruth(beamExpansions, beamSize);
+}
+
+void testCrossEntropyOverBeam(bool useGpu,
+                              size_t beamSize,
+                              vector<SingleBeamExpansion>& beams) {
+  TestConfig config;
+  config.layerConfig.set_type("cross_entropy_over_beam");
+
+  size_t seqNum = 0;
+  for (size_t i = 1; i < beams.size(); ++i) {
+    const SingleBeamExpansion& beam = beams[i];
+    // create scores for all the candidates
+    MatrixPtr candidateScorePtr =
+        Matrix::create(beam.candidateScores.size(), 1, false, false);
+    candidateScorePtr->copyFrom(beam.candidateScores.data(),
+                                beam.candidateScores.size());
+
+    ostringstream paramName;
+    paramName << "candidate_scores_" << i;
+
+    if (beam.subSeqStartPos.size() > 1) {
+      seqNum = beam.subSeqStartPos.size() - 1;
+      config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA,
+                                  paramName.str(),
+                                  candidateScorePtr,
+                                  beam.seqStartPos,
+                                  beam.subSeqStartPos});
+    } else {
+      seqNum = beam.seqStartPos.size() - 1;
+      config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA,
+                                  paramName.str(),
+                                  candidateScorePtr,
+                                  beam.seqStartPos});
+    }
+    config.layerConfig.add_inputs();
+
+    // create indices for the selected candidates
+    MatrixPtr selectedCandidates =
+        Matrix::create(seqNum, beamSize, false, false);
+    selectedCandidates->copyFrom(beam.selectedIndices.data(),
+                                 beam.selectedIndices.size());
+    paramName.clear();
+    paramName << "selected_candidates_" << i;
+    config.inputDefs.push_back(
+        {INPUT_SELF_DEFINE_DATA, paramName.str(), selectedCandidates});
+    config.layerConfig.add_inputs();
+
+    // create the ground truth
+    paramName.clear();
+    paramName << "label_" << i;
+    config.inputDefs.push_back(
+        {INPUT_SELF_DEFINE_DATA, paramName.str(), beam.groundTruth});
+    config.layerConfig.add_inputs();
+  }
+
+  testLayerGrad(
+      config, "cross_entropy_over_beam", seqNum, false, useGpu, false);
+}
+
+TEST(Layer, CrossEntropyOverBeam) {
+  LOG(INFO) << "SEED = " << SEED;
+  const size_t beamSize = 1 + rand() % MAX_BEAM_SIZE;
+  LOG(INFO) << "beamSize = " << beamSize;
+
+  // TODO(caoying): test with random beam expansions.
+  const size_t expansionCount = 3;
+  vector<SingleBeamExpansion> beams;
+  genRandomBeamExpansion(expansionCount, beamSize, beams);
+
+  for (bool useGpu : {false, true})
+    testCrossEntropyOverBeam(useGpu, beamSize, beams);
+}
+
+int main(int argc, char** argv) {
+  initMain(argc, argv);
+  hl_start();
+  hl_init(FLAGS_gpu_id);
+  FLAGS_thread_local_rand_use_global_seed = true;
+  srand(SEED);
+  testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}

paddle/gserver/tests/test_LayerGrad.cpp

@@ -1246,6 +1246,75 @@ TEST(Layer, PoolLayer) {
 #endif
 }
 
+void setPool3DConfig(TestConfig* config,
+                     PoolConfig* pool,
+                     const string& poolType) {
+  // filter size
+  const int NUM_FILTERS = 16;
+  const int FILTER_SIZE = 3;
+  const int FILTER_SIZE_Y = 3;
+  const int FILTER_SIZE_Z = 3;
+  const int CHANNELS = 16;
+
+  (*config).biasSize = 0;
+  (*config).layerConfig.set_type("pool3d");
+  (*config).layerConfig.set_num_filters(NUM_FILTERS);
+
+  int kw = FILTER_SIZE, kh = FILTER_SIZE_Y, kd = FILTER_SIZE_Z;
+  int pw = 0, ph = 0, pd = 0;
+  int sw = 2, sh = 2, sd = 2;
+
+  pool->set_pool_type(poolType);
+  pool->set_pool_type("avg");
+  pool->set_channels(CHANNELS);
+  pool->set_size_x(kw);
+  pool->set_size_y(kh);
+  pool->set_size_z(kd);
+  pool->set_padding(0);
+  pool->set_padding_y(0);
+  pool->set_padding_z(0);
+  pool->set_stride(sw);
+  pool->set_stride_y(sh);
+  pool->set_stride_z(sd);
+  pool->set_start(0);
+  int ow = outputSize(pool->img_size(), kw, pw, sw, /* caffeMode */ false);
+  int oh = outputSize(pool->img_size_y(), kh, ph, sh, /* caffeMode */ false);
+  int od = outputSize(pool->img_size_z(), kd, pd, sd, /* caffeMode */ false);
+  pool->set_output_x(ow);
+  pool->set_output_y(oh);
+  pool->set_output_z(od);
+}
+
+void testPool3DLayer(const string& poolType, bool trans, bool useGpu) {
+  TestConfig config;
+  config.inputDefs.push_back({INPUT_DATA, "layer_0", 11664, 0});
+  LayerInputConfig* input = config.layerConfig.add_inputs();
+  PoolConfig* pool = input->mutable_pool_conf();
+
+  const int IMAGE_SIZE = 9;
+  const int IMAGE_SIZE_Y = 9;
+  const int IMAGE_SIZE_Z = 9;
+
+  pool->set_img_size(IMAGE_SIZE);
+  pool->set_img_size_y(IMAGE_SIZE_Y);
+  pool->set_img_size_z(IMAGE_SIZE_Z);
+
+  setPool3DConfig(&config, pool, poolType);
+  config.layerConfig.set_size(pool->output_x() * pool->output_y() *
+                              pool->channels());
+
+  testLayerGrad(config, "pool3d", 100, trans, useGpu);
+}
+
+TEST(Layer, Pool3DLayer) {
+  testPool3DLayer("avg", /* trans= */ false, /* useGpu= */ false);
+  testPool3DLayer("max", /* trans= */ false, /* useGpu= */ false);
+#ifndef PADDLE_ONLY_CPU
+  testPool3DLayer("avg", /* trans= */ false, /* useGpu= */ true);
+  testPool3DLayer("max", /* trans= */ false, /* useGpu= */ true);
+#endif
+}
+
 void testSppLayer(const string& poolType,
                   const int pyramidHeight,
                   bool trans,
@@ -2047,6 +2116,159 @@ TEST(Layer, RowL2NormLayer) {
   }
 }
 
+void test3DConvLayer(const string& type, bool trans, bool useGpu) {
+  // filter size
+  const int NUM_FILTERS = 6;
+  // const int CHANNELS = 3;
+  const int FILTER_SIZE = 3;
+  const int FILTER_SIZE_Y = 3;
+  const int FILTER_SIZE_Z = 3;
+
+  // input image
+  const int CHANNELS = 3;
+  const int IMAGE_SIZE = 9;
+  const int IMAGE_SIZE_Y = 9;
+  const int IMAGE_SIZE_Z = 9;
+
+  TestConfig config;
+  config.biasSize = NUM_FILTERS;
+  config.layerConfig.set_type(type);
+  config.layerConfig.set_num_filters(NUM_FILTERS);
+  config.layerConfig.set_partial_sum(1);
+  config.layerConfig.set_shared_biases(true);
+
+  // Setting up conv3D-trans layer
+  LayerInputConfig* input = config.layerConfig.add_inputs();
+  ConvConfig* conv = input->mutable_conv_conf();
+
+  conv->set_channels(CHANNELS);
+  conv->set_filter_size(FILTER_SIZE);
+  conv->set_filter_size_y(FILTER_SIZE_Y);
+  conv->set_filter_size_z(FILTER_SIZE_Z);
+  conv->set_padding(0);
+  conv->set_padding_y(0);
+  conv->set_padding_z(0);
+  conv->set_stride(2);
+  conv->set_stride_y(2);
+  conv->set_stride_z(2);
+  conv->set_img_size(IMAGE_SIZE);
+  conv->set_img_size_y(IMAGE_SIZE_Y);
+  conv->set_img_size_z(IMAGE_SIZE_Z);
+  conv->set_output_x(outputSize(conv->img_size(),
+                                conv->filter_size(),
+                                conv->padding(),
+                                conv->stride(),
+                                /*  caffeMode */ true));
+  conv->set_output_y(outputSize(conv->img_size_y(),
+                                conv->filter_size_y(),
+                                conv->padding_y(),
+                                conv->stride_y(),
+                                /*  caffeMode */ true));
+  conv->set_output_z(outputSize(conv->img_size_z(),
+                                conv->filter_size_z(),
+                                conv->padding_z(),
+                                conv->stride_z(),
+                                /*  caffeMode */ true));
+
+  config.layerConfig.set_size(conv->output_x() * conv->output_y() *
+                              conv->output_z() * NUM_FILTERS);
+  conv->set_groups(1);
+  conv->set_filter_channels(conv->channels() / conv->groups());
+  config.inputDefs.push_back(
+      {INPUT_DATA,
+       "layer_0",
+       CHANNELS * IMAGE_SIZE * IMAGE_SIZE_Y * IMAGE_SIZE_Z,
+       conv->filter_channels() * FILTER_SIZE * FILTER_SIZE_Y * FILTER_SIZE_Z *
+           NUM_FILTERS});
+
+  testLayerGrad(config, "conv3D", 10, trans, useGpu);
+  // Use small batch_size and useWeight=true to test biasGrad
+  testLayerGrad(config, "conv3D", 2, trans, useGpu, true, 0.02);
+}
+
+TEST(Layer, test3DConvLayer) {
+  test3DConvLayer("conv3d", /* trans= */ false, /* useGpu= */ false);
+#ifndef PADDLE_ONLY_CPU
+  test3DConvLayer("conv3d", /* trans= */ false, /* useGpu= */ true);
+#endif
+}
+
+void test3DDeConvLayer(const string& type, bool trans, bool useGpu) {
+  // filter size
+  const int NUM_FILTERS = 6;
+  // const int CHANNELS = 3;
+  const int FILTER_SIZE = 3;
+  const int FILTER_SIZE_Y = 3;
+  const int FILTER_SIZE_Z = 3;
+
+  // input image
+  const int CHANNELS = 3;
+  const int IMAGE_SIZE = 4;
+  const int IMAGE_SIZE_Y = 6;
+  const int IMAGE_SIZE_Z = 6;
+
+  // Setting up conv-trans layer
+  TestConfig config;
+  config.biasSize = NUM_FILTERS;
+  config.layerConfig.set_type("deconv3d");
+  config.layerConfig.set_num_filters(NUM_FILTERS);
+  config.layerConfig.set_partial_sum(1);
+  config.layerConfig.set_shared_biases(true);
+
+  LayerInputConfig* input = config.layerConfig.add_inputs();
+  ConvConfig* conv = input->mutable_conv_conf();
+
+  conv->set_channels(CHANNELS);
+  conv->set_filter_size(FILTER_SIZE);
+  conv->set_filter_size_y(FILTER_SIZE_Y);
+  conv->set_filter_size_z(FILTER_SIZE_Z);
+  conv->set_padding(0);
+  conv->set_padding_y(0);
+  conv->set_padding_z(0);
+  conv->set_stride(2);
+  conv->set_stride_y(2);
+  conv->set_stride_z(2);
+  conv->set_img_size(IMAGE_SIZE);
+  conv->set_img_size_y(IMAGE_SIZE_Y);
+  conv->set_img_size_z(IMAGE_SIZE_Z);
+  conv->set_output_x(imageSize(conv->img_size(),
+                               conv->filter_size(),
+                               conv->padding(),
+                               conv->stride(),
+                               true));
+  conv->set_output_y(imageSize(conv->img_size_y(),
+                               conv->filter_size_y(),
+                               conv->padding_y(),
+                               conv->stride_y(),
+                               true));
+  conv->set_output_z(imageSize(conv->img_size_z(),
+                               conv->filter_size_z(),
+                               conv->padding_z(),
+                               conv->stride_z(),
+                               true));
+  config.layerConfig.set_size(conv->output_x() * conv->output_y() *
+                              conv->output_z() * NUM_FILTERS);
+  conv->set_groups(1);
+  conv->set_filter_channels(conv->channels() / conv->groups());
+  config.inputDefs.push_back(
+      {INPUT_DATA,
+       "layer_0",
+       CHANNELS * IMAGE_SIZE * IMAGE_SIZE_Y * IMAGE_SIZE_Z,
+       conv->filter_channels() * FILTER_SIZE * FILTER_SIZE_Y * FILTER_SIZE_Z *
+           NUM_FILTERS});
+
+  testLayerGrad(config, "deconv3D", 10, trans, useGpu);
+  // Use small batch_size and useWeight=true to test biasGrad
+  testLayerGrad(config, "deconv3D", 2, trans, useGpu, true, 0.02);
+}
+
+TEST(Layer, test3DDeConvLayer) {
+  test3DDeConvLayer("deconv3d", /* trans= */ false, /* useGpu= */ false);
+#ifndef PADDLE_ONLY_CPU
+  test3DDeConvLayer("deconv3d", /* trans= */ false, /* useGpu= */ true);
+#endif
+}
+
 TEST(Layer, ScaleShiftLayer) {
   const size_t batchSize = 16;
   const size_t size = 32;

paddle/math/Allocator.h

@@ -48,7 +48,13 @@ public:
    */
   virtual void* alloc(size_t size) {
     void* ptr;
+#ifdef PADDLE_USE_MKLDNN
+    // refer to https://github.com/01org/mkl-dnn/blob/master/include/mkldnn.hpp
+    // memory alignment
+    CHECK_EQ(posix_memalign(&ptr, 4096ul, size), 0);
+#else
     CHECK_EQ(posix_memalign(&ptr, 32ul, size), 0);
+#endif
     CHECK(ptr) << "Fail to allocate CPU memory: size=" << size;
     return ptr;
   }

paddle/math/CMakeLists.txt

@@ -14,6 +14,17 @@
 #
 file(GLOB MATH_HEADERS . *.h)
 file(GLOB MATH_SOURCES . *.cpp)
+
+if(NOT WITH_MKLDNN)
+    set(DNN_HEADER "${CMAKE_CURRENT_SOURCE_DIR}/MKLDNNMatrix.h")
+    set(DNN_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/MKLDNNMatrix.cpp")
+    list(REMOVE_ITEM MATH_HEADERS "${DNN_HEADER}")
+    list(REMOVE_ITEM MATH_SOURCES "${DNN_SOURCE}")
+    message(STATUS "Skip compiling with MKLDNNMatrix")
+else()
+    message(STATUS "Compile with MKLDNNMatrix")
+endif()
+
 set(MATH_SOURCES
     "${PADDLE_SOURCE_DIR}/paddle/math/BaseMatrix.cu"
     "${PADDLE_SOURCE_DIR}/paddle/math/TrainingAlgorithmOp.cu"

paddle/math/MKLDNNMatrix.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 "MKLDNNMatrix.h"
+
+using namespace mkldnn;  // NOLINT
+
+namespace paddle {
+
+MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m, memory::primitive_desc pd) {
+  memory::desc md = pd.desc();
+  size_t ndims = md.data.ndims;
+  int* dims = md.data.dims;
+  CHECK(ndims > 0) << "Input dims should not be empty";
+  size_t cnts = 1;
+  for (size_t i = 0; i < ndims; ++i) {
+    cnts *= dims[i];
+  }
+
+  if (m == nullptr) {
+    size_t height = dims[0];
+    size_t width = cnts / dims[0];
+    m = Matrix::create(height, width, false, false);
+  }
+
+  CHECK(m) << " Matrix should not be empty";
+  CpuMatrixPtr cpuMatrix = std::dynamic_pointer_cast<CpuMatrix>(m);
+  CHECK(cpuMatrix) << "Only support create from CPU matrix yet";
+
+  CHECK_EQ(cnts, m->getElementCnt()) << "Count size does not match";
+  return std::make_shared<MKLDNNMatrix>(
+      m->getData(), m->getHeight(), m->getWidth(), pd);
+}
+
+MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m,
+                                     memory::dims dims,
+                                     memory::format fmt,
+                                     engine& eg,
+                                     mkldnn::memory::data_type dtype) {
+  return create(m, memory::primitive_desc(memory::desc(dims, dtype, fmt), eg));
+}
+
+void MKLDNNMatrix::reorderDataFrom(const MKLDNNMatrixPtr& m,
+                                   memory::format srcFmt,
+                                   memory::dims targetDim) {
+  memory::format dstFmt = getFormat();
+  if (srcFmt == dstFmt) {
+    return;
+  }
+  CHECK_EQ(getElementCnt(), m->getElementCnt()) << "size should equal";
+  reorderOnce(getData(), m->getData(), srcFmt, dstFmt, targetDim);
+}
+
+void MKLDNNMatrix::reorderDataTo(const MKLDNNMatrixPtr& m,
+                                 memory::format dstFmt,
+                                 memory::dims targetDim) {
+  memory::format srcFmt = getFormat();
+  if (srcFmt == dstFmt) {
+    return;
+  }
+  CHECK_EQ(getElementCnt(), m->getElementCnt()) << "size should equal";
+  reorderOnce(getData(), m->getData(), srcFmt, dstFmt, targetDim);
+}
+
+void MKLDNNMatrix::reorderOnce(void* srcData,
+                               void* dstData,
+                               memory::format srcFmt,
+                               memory::format dstFmt,
+                               memory::dims dm) {
+  CHECK(srcData);
+  CHECK(dstData);
+  MatrixPtr tmpSrc;
+  if (dstData == srcData) {
+    // inplace data
+    size_t sz = 1;
+    for (size_t i = 0; i < dm.size(); ++i) {
+      sz *= dm[i];
+    }
+    tmpSrc = Matrix::create(sz, 1, false, false);
+    tmpSrc->copyFrom((real*)srcData, sz);
+    srcData = tmpSrc->getData();
+  }
+
+  auto dtype = this->getDtype();
+  auto srcMD = memory::desc(dm, dtype, srcFmt);
+  auto dstMD = memory::desc(dm, dtype, dstFmt);
+
+  auto eg = this->getEngine();
+  auto src = memory(memory::primitive_desc(srcMD, eg), srcData);
+  auto dst = memory(memory::primitive_desc(dstMD, eg), dstData);
+
+  auto r = reorder(src, dst);
+  stream(stream::kind::eager).submit({r}).wait();
+}
+
+void MKLDNNMatrix::downSpatial() {
+  int fmt = getFormat();
+  if (!(fmt == memory::format::nchw || fmt == memory::format::oihw)) {
+    // only support nchw and oihw yet, later can support more like nhwc, ihwo
+    return;
+  }
+
+  // TODO(TJ): change H(height) and W(width) if support nhwc or more
+  const int H = 2, W = 3;
+  memory::dims srcDims = getDims();
+  if (srcDims[H] != 1 || srcDims[W] != 1) {
+    // can not down spatial
+    return;
+  }
+
+  memory::dims dstDims = memory::dims{srcDims[0], srcDims[1]};
+  memory::format dstFmt;
+  switch (fmt) {
+    case memory::format::nchw:
+      dstFmt = memory::format::nc;
+      break;
+    case memory::format::oihw:
+      dstFmt = memory::format::oi;
+      break;
+    default:
+      LOG(FATAL) << "unsupported format";
+  }
+  memory::desc md = memory::desc(dstDims, getDtype(), dstFmt);
+  memory::primitive_desc pd = memory::primitive_desc(md, getEngine());
+  mkldnn_primitive_t result;
+  mkldnn::error::wrap_c_api(
+      mkldnn_primitive_create(&result, pd.get(), nullptr, nullptr),
+      "could not create a memory primitive");
+  reset(result);
+  set_data_handle(getData());
+}
+
+}  // namespace paddle

paddle/math/MKLDNNMatrix.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 <vector>
+#include "Matrix.h"
+#include "mkldnn.hpp"
+#include "paddle/parameter/Parameter.h"
+
+namespace paddle {
+
+class MKLDNNMatrix;
+typedef std::shared_ptr<MKLDNNMatrix> MKLDNNMatrixPtr;
+
+/**
+ * @brief MKLDNN Matrix.
+ *
+ */
+class MKLDNNMatrix : public CpuMatrix, public mkldnn::memory {
+public:
+  MKLDNNMatrix(real* data,
+               size_t height,
+               size_t width,
+               mkldnn::memory::primitive_desc pd)
+      : CpuMatrix(data, height, width, false), mkldnn::memory(pd, data) {}
+
+  ~MKLDNNMatrix() {}
+
+  /**
+   * Create MKLDNNMatrix from a MatrixPtr and memory primitive_desc
+   */
+  static MKLDNNMatrixPtr create(MatrixPtr m, mkldnn::memory::primitive_desc pd);
+
+  /**
+   * Create MKLDNNMatrix from a MatrixPtr and memory details info
+   */
+  static MKLDNNMatrixPtr create(
+      MatrixPtr m,
+      mkldnn::memory::dims dims,
+      mkldnn::memory::format fmt,
+      mkldnn::engine& eg,
+      mkldnn::memory::data_type dtype = mkldnn::memory::data_type::f32);
+
+public:
+  /**
+   * Reorder this MKLDNNMatrix from other format.
+   * Support inplace reorder.
+   * @note: this function would only reorder the data layout.
+   *        will NOT change this original dim or format info
+   */
+  void reorderDataFrom(const MKLDNNMatrixPtr& m,
+                       memory::format srcFmt,
+                       memory::dims targetDim);
+
+  /**
+   * Reorder this MKLDNNMatrix to other format.
+   * Support inplace reorder.
+   * @note: this function would only reorder the data layout.
+   *        will NOT change the dst dim or format info
+   */
+  void reorderDataTo(const MKLDNNMatrixPtr& m,
+                     memory::format dstFmt,
+                     memory::dims targetDim);
+
+  /**
+   * Dimensionality reduction.
+   * Change format "nchw --> nc" or "oihw --> oi" if the h and w are both 1
+   */
+  void downSpatial();
+
+  /**
+   * Update the memory data handle.
+   * Caution: This will not check the buffer size of the data,
+   *          it should be coverd by user.
+   */
+  void updateData(void* data) { set_data_handle(data); }
+
+  /**
+   * Get primitive descriptor.
+   */
+  mkldnn::memory::primitive_desc getPrimitiveDesc() {
+    return this->get_primitive_desc();
+  }
+
+  /**
+   * Get memory descriptor.
+   */
+  mkldnn::memory::desc getMemoryDesc() { return getPrimitiveDesc().desc(); }
+
+  /**
+   * Get dimensions.
+   */
+  mkldnn::memory::dims getDims() {
+    mkldnn::memory::desc md = getMemoryDesc();
+    const int* src = md.data.dims;
+    int ndims = md.data.ndims;
+    mkldnn::memory::dims dst;
+    dst.resize(ndims);
+    for (int i = 0; i < ndims; ++i) {
+      dst[i] = src[i];
+    }
+    return dst;
+  }
+
+  /**
+   * Get format.
+   */
+  mkldnn::memory::format getFormat() {
+    return (mkldnn::memory::format)(getMemoryDesc().data.format);
+  }
+
+  /**
+   * Get memory data type.
+   */
+  mkldnn::memory::data_type getDtype() {
+    return (mkldnn::memory::data_type)(getMemoryDesc().data.data_type);
+  }
+
+  /**
+   * Get engine.
+   */
+  mkldnn::engine getEngine() { return getPrimitiveDesc().get_engine(); }
+
+protected:
+  /**
+   * Do reorder once.
+   * Can support inplace.
+   */
+  void reorderOnce(void* srcData,
+                   void* dstData,
+                   memory::format srcFmt,
+                   memory::format dstFmt,
+                   memory::dims dm);
+};
+
+}  // namespace paddle

paddle/math/Matrix.cpp

@@ -1190,6 +1190,221 @@ void GpuMatrix::avgPoolBackward(Matrix& outGrad,
                       outGrad.getStride());
 }
 
+void GpuMatrix::maxPool3DForward(Matrix& inputMat,
+                                 Matrix& maxPoolIdx,
+                                 size_t channels,
+                                 size_t imgSizeD,
+                                 size_t imgSizeH,
+                                 size_t imgSizeW,
+                                 size_t outputD,
+                                 size_t outputH,
+                                 size_t outputW,
+                                 size_t sizeZ,
+                                 size_t sizeY,
+                                 size_t sizeX,
+                                 size_t strideD,
+                                 size_t strideH,
+                                 size_t strideW,
+                                 size_t paddingD,
+                                 size_t paddingH,
+                                 size_t paddingW) {
+  CHECK(inputMat.useGpu_) << "Matrix type are not correct";
+
+  real* inputData = inputMat.getData();
+  real* maxPoolIdxData = maxPoolIdx.getData();
+  size_t num = inputMat.getHeight();
+  size_t width = imgSizeW;
+  size_t height = imgSizeH;
+  size_t depth = imgSizeD;
+  CHECK(depth * height * width * channels == inputMat.getWidth());
+  CHECK(height_ == inputMat.getHeight());
+  CHECK(width_ == outputD * outputH * outputW * channels);
+
+  hl_maxpool3D_forward(num,
+                       inputData,
+                       channels,
+                       depth,
+                       height,
+                       width,
+                       outputD,
+                       outputH,
+                       outputW,
+                       sizeZ,
+                       sizeY,
+                       sizeX,
+                       strideD,
+                       strideH,
+                       strideW,
+                       paddingD,
+                       paddingH,
+                       paddingW,
+                       getData(),
+                       maxPoolIdxData,
+                       getStride());
+}
+
+void GpuMatrix::maxPool3DBackward(Matrix& outGrad,
+                                  Matrix& maxPoolIdx,
+                                  size_t imgSizeD,
+                                  size_t imgSizeH,
+                                  size_t imgSizeW,
+                                  size_t outputD,
+                                  size_t outputH,
+                                  size_t outputW,
+                                  size_t sizeZ,
+                                  size_t sizeY,
+                                  size_t sizeX,
+                                  size_t strideD,
+                                  size_t strideH,
+                                  size_t strideW,
+                                  size_t paddingD,
+                                  size_t paddingH,
+                                  size_t paddingW,
+                                  real scaleTargets,
+                                  real scaleOutput) {
+  CHECK(outGrad.useGpu_ && maxPoolIdx.useGpu_) << "Matrix type are not equal";
+
+  real* outDiff = outGrad.getData();
+  real* maxPoolIdxData = maxPoolIdx.getData();
+  size_t frameNum = getHeight();
+  size_t channels = outGrad.getWidth() / outputD / outputH / outputW;
+  size_t width = imgSizeW;
+  size_t height = imgSizeH;
+  size_t depth = imgSizeD;
+  CHECK(depth * height * width * channels == getWidth());
+  CHECK(width_ == depth * width * height * channels);
+  CHECK(outGrad.getHeight() == maxPoolIdx.getHeight() &&
+        outGrad.getWidth() == maxPoolIdx.getWidth());
+
+  hl_maxpool3D_backward(frameNum,
+                        outDiff,
+                        channels,
+                        depth,
+                        height,
+                        width,
+                        outputD,
+                        outputH,
+                        outputW,
+                        sizeZ,
+                        sizeY,
+                        sizeX,
+                        strideD,
+                        strideH,
+                        strideW,
+                        paddingD,
+                        paddingH,
+                        paddingW,
+                        scaleTargets,
+                        scaleOutput,
+                        getData(),
+                        maxPoolIdxData,
+                        outGrad.getStride());
+}
+
+void GpuMatrix::avgPool3DForward(Matrix& inputMat,
+                                 size_t channels,
+                                 size_t imgSizeD,
+                                 size_t imgSizeH,
+                                 size_t imgSizeW,
+                                 size_t outputD,
+                                 size_t outputH,
+                                 size_t outputW,
+                                 size_t sizeZ,
+                                 size_t sizeY,
+                                 size_t sizeX,
+                                 size_t strideD,
+                                 size_t strideH,
+                                 size_t strideW,
+                                 size_t paddingD,
+                                 size_t paddingH,
+                                 size_t paddingW) {
+  CHECK(inputMat.useGpu_) << "Matrix type are not equal";
+
+  real* inputData = inputMat.getData();
+  size_t frameNum = inputMat.getHeight();
+  size_t height = imgSizeH;
+  size_t width = imgSizeW;
+  size_t depth = imgSizeD;
+  CHECK(depth * height * width * channels == inputMat.getWidth());
+  CHECK(height_ == inputMat.getHeight());
+  CHECK(width_ == outputD * outputH * outputW * channels);
+
+  hl_avgpool3D_forward(frameNum,
+                       inputData,
+                       channels,
+                       depth,
+                       height,
+                       width,
+                       outputD,
+                       outputH,
+                       outputW,
+                       sizeZ,
+                       sizeY,
+                       sizeX,
+                       strideD,
+                       strideH,
+                       strideW,
+                       paddingD,
+                       paddingH,
+                       paddingW,
+                       getData(),
+                       getStride());
+}
+
+void GpuMatrix::avgPool3DBackward(Matrix& outGrad,
+                                  size_t imgSizeD,
+                                  size_t imgSizeH,
+                                  size_t imgSizeW,
+                                  size_t outputD,
+                                  size_t outputH,
+                                  size_t outputW,
+                                  size_t sizeZ,
+                                  size_t sizeY,
+                                  size_t sizeX,
+                                  size_t strideD,
+                                  size_t strideH,
+                                  size_t strideW,
+                                  size_t paddingD,
+                                  size_t paddingH,
+                                  size_t paddingW,
+                                  real scaleTargets,
+                                  real scaleOutput) {
+  CHECK(outGrad.useGpu_) << "Matrix type are not equal";
+
+  real* outDiff = outGrad.getData();
+  size_t frameNum = outGrad.getHeight();
+  size_t channels = outGrad.getWidth() / outputD / outputH / outputW;
+  size_t height = imgSizeH;
+  size_t width = imgSizeW;
+  size_t depth = imgSizeD;
+  CHECK(depth * height * width * channels == width_);
+  CHECK(height_ == outGrad.getHeight());
+  CHECK(outGrad.getWidth() == outputD * outputH * outputW * channels);
+
+  hl_avgpool3D_backward(frameNum,
+                        outDiff,
+                        channels,
+                        depth,
+                        height,
+                        width,
+                        outputD,
+                        outputH,
+                        outputW,
+                        sizeZ,
+                        sizeY,
+                        sizeX,
+                        strideD,
+                        strideH,
+                        strideW,
+                        paddingD,
+                        paddingH,
+                        paddingW,
+                        scaleTargets,
+                        scaleOutput,
+                        getData(),
+                        outGrad.getStride());
+}
+
 void GpuMatrix::maxSequenceForward(Matrix& input,
                                    const IVector& sequence,
                                    IVector& index) {
@@ -1389,6 +1604,72 @@ void GpuMatrix::multiBinaryLabelCrossEntropyBp(Matrix& output, Matrix& label) {
       output_d, grad_d, mat_d, height_, width_);
 }
 
+void GpuMatrix::vol2Col(real* dataSrc,
+                        int channels,
+                        int depth,
+                        int height,
+                        int width,
+                        int filterD,
+                        int filterH,
+                        int filterW,
+                        int strideD,
+                        int strideH,
+                        int strideW,
+                        int paddingD,
+                        int paddingH,
+                        int paddingW) {
+  hl_matrix_vol2Col(dataSrc,
+                    channels,
+                    depth,
+                    height,
+                    width,
+                    filterD,
+                    filterH,
+                    filterW,
+                    strideD,
+                    strideH,
+                    strideW,
+                    paddingD,
+                    paddingH,
+                    paddingW,
+                    getData());
+}
+
+void GpuMatrix::col2Vol(real* dataDst,
+                        int channels,
+                        int depth,
+                        int height,
+                        int width,
+                        int filterD,
+                        int filterH,
+                        int filterW,
+                        int strideD,
+                        int strideH,
+                        int strideW,
+                        int paddingD,
+                        int paddingH,
+                        int paddingW,
+                        real alpha,
+                        real beta) {
+  hl_matrix_col2Vol(dataDst,
+                    channels,
+                    depth,
+                    height,
+                    width,
+                    filterD,
+                    filterH,
+                    filterW,
+                    strideD,
+                    strideH,
+                    strideW,
+                    paddingD,
+                    paddingH,
+                    paddingW,
+                    getData(),
+                    alpha,
+                    beta);
+}
+
 /**
  * CpuMatrix
  */
@@ -1930,6 +2211,276 @@ void CpuMatrix::avgPoolBackward(Matrix& input,
   }
 }
 
+void CpuMatrix::maxPool3DForward(Matrix& inputMat,
+                                 Matrix& maxPoolIdx,
+                                 size_t channels,
+                                 size_t imgSizeD,
+                                 size_t imgSizeH,
+                                 size_t imgSizeW,
+                                 size_t outputD,
+                                 size_t outputH,
+                                 size_t outputW,
+                                 size_t sizeZ,
+                                 size_t sizeY,
+                                 size_t sizeX,
+                                 size_t strideD,
+                                 size_t strideH,
+                                 size_t strideW,
+                                 size_t paddingD,
+                                 size_t paddingH,
+                                 size_t paddingW) {
+  real* inputData = inputMat.getData();
+  real* outData = getData();
+  real* maxPoolIdxData = maxPoolIdx.getData();
+  size_t num = inputMat.getHeight();
+  size_t inWidth = imgSizeW;
+  size_t inHeight = imgSizeH;
+  size_t inDepth = imgSizeD;
+  CHECK(inHeight * inWidth * inDepth == inputMat.getWidth() / channels);
+  CHECK_EQ(num, this->getHeight());
+  CHECK_EQ(channels * outputH * outputW * outputD, this->getWidth());
+  size_t outStride = getStride();
+
+  /* initialize the data_ */
+  for (size_t i = 0; i < height_; i++) {
+    for (size_t j = 0; j < width_; j++) {
+      outData[(i)*outStride + j] = -(real)FLT_MAX;
+      maxPoolIdxData[(i)*outStride + j] = -1;
+    }
+  }
+
+  /* pool max one by one */
+  for (size_t n = 0; n < num; ++n) {  // frame by frame
+    if (!isContiguous()) {
+      outData = getData() + n * outStride;
+      maxPoolIdxData = maxPoolIdx.getData() + n * outStride;
+    }
+    for (size_t c = 0; c < channels; ++c) {  // channel by channel
+      for (size_t pd = 0; pd < outputD; ++pd) {
+        for (size_t ph = 0; ph < outputH; ++ph) {
+          for (size_t pw = 0; pw < outputW; ++pw) {
+            int dstart = pd * strideD - paddingD;
+            int hstart = ph * strideH - paddingH;
+            int wstart = pw * strideW - paddingW;
+            int dend = std::min(dstart + sizeZ, inDepth);
+            int hend = std::min(hstart + sizeY, inHeight);
+            int wend = std::min(wstart + sizeX, inWidth);
+            dstart = std::max(dstart, 0);
+            hstart = std::max(hstart, 0);
+            wstart = std::max(wstart, 0);
+            int maxIdx = -1;
+            real maxOutData = outData[(pd * outputH + ph) * outputW + pw];
+            for (int d = dstart; d < dend; ++d) {
+              for (int h = hstart; h < hend; ++h) {
+                for (int w = wstart; w < wend; ++w) {
+                  if (maxOutData <
+                      inputData[(d * inHeight + h) * inWidth + w]) {
+                    maxOutData = inputData[(d * inHeight + h) * inWidth + w];
+                    maxIdx = (d * inHeight + h) * inWidth + w;
+                  }
+                }
+              }
+            }
+            outData[(pd * outputH + ph) * outputW + pw] = maxOutData;
+            maxPoolIdxData[(pd * outputH + ph) * outputW + pw] = maxIdx;
+          }
+        }
+      }
+      // compute offset
+      inputData += inDepth * inHeight * inWidth;
+      outData += outputD * outputH * outputW;
+      maxPoolIdxData += outputD * outputH * outputW;
+    }
+  }
+}
+
+void CpuMatrix::maxPool3DBackward(Matrix& outGrad,
+                                  Matrix& maxPoolIdx,
+                                  size_t imgSizeD,
+                                  size_t imgSizeH,
+                                  size_t imgSizeW,
+                                  size_t outputD,
+                                  size_t outputH,
+                                  size_t outputW,
+                                  size_t sizeZ,
+                                  size_t sizeY,
+                                  size_t sizeX,
+                                  size_t strideD,
+                                  size_t strideH,
+                                  size_t strideW,
+                                  size_t paddingD,
+                                  size_t paddingH,
+                                  size_t paddingW,
+                                  real scaleTargets,
+                                  real scaleOutput) {
+  size_t num = getHeight();
+  size_t channels = size_t(width_ / imgSizeD / imgSizeH / imgSizeW);
+  CHECK(maxPoolIdx.getHeight() == outGrad.getHeight() &&
+        maxPoolIdx.getWidth() == outGrad.getWidth());
+
+  real* tgtGrad = getData();
+  real* otGrad = outGrad.getData();
+  real* maxPoolIdxData = maxPoolIdx.getData();
+  size_t outStride = outGrad.getStride();
+
+  for (size_t n = 0; n < num; ++n) {
+    if (!outGrad.isContiguous()) {
+      otGrad = outGrad.getData() + n * outStride;
+      maxPoolIdxData = maxPoolIdx.getData() + n * outStride;
+    }
+    for (size_t c = 0; c < channels; ++c) {
+      for (size_t pd = 0; pd < outputD; ++pd) {
+        for (size_t ph = 0; ph < outputH; ++ph) {
+          for (size_t pw = 0; pw < outputW; ++pw) {
+            const size_t index = (pd * outputH + ph) * outputW + pw;
+            const size_t tgtIdx = static_cast<size_t>(maxPoolIdxData[index]);
+            tgtGrad[tgtIdx] =
+                scaleTargets * tgtGrad[tgtIdx] + scaleOutput * otGrad[index];
+          }
+        }
+      }
+      // offset
+      tgtGrad += imgSizeD * imgSizeH * imgSizeW;
+      otGrad += outputD * outputH * outputW;
+      maxPoolIdxData += outputD * outputH * outputW;
+    }
+  }
+}
+
+void CpuMatrix::avgPool3DForward(Matrix& input,
+                                 size_t channels,
+                                 size_t imgSizeD,
+                                 size_t imgSizeH,
+                                 size_t imgSizeW,
+                                 size_t outputD,
+                                 size_t outputH,
+                                 size_t outputW,
+                                 size_t sizeZ,
+                                 size_t sizeY,
+                                 size_t sizeX,
+                                 size_t strideD,
+                                 size_t strideH,
+                                 size_t strideW,
+                                 size_t paddingD,
+                                 size_t paddingH,
+                                 size_t paddingW) {
+  // The main loop
+  size_t num = input.getHeight();
+  size_t inDepth = imgSizeD;
+  size_t inHeight = imgSizeH;
+  size_t inWidth = imgSizeW;
+  CHECK(inDepth * inHeight * inWidth * channels == input.getWidth());
+  CHECK(outputD * outputH * outputW * channels * num == height_ * width_);
+  real* tgtData = getData();
+  real* inData = input.getData();
+
+  for (size_t n = 0; n < num; ++n) {
+    if (!isContiguous()) {
+      tgtData = data_ + n * getStride();
+    }
+    for (size_t c = 0; c < channels; ++c) {
+      for (size_t pd = 0; pd < outputD; ++pd) {
+        for (size_t ph = 0; ph < outputH; ++ph) {
+          for (size_t pw = 0; pw < outputW; ++pw) {
+            int dstart = pd * strideD - paddingD;
+            int hstart = ph * strideH - paddingH;
+            int wstart = pw * strideW - paddingW;
+            int dend = std::min(dstart + sizeZ, inDepth + paddingD);
+            int hend = std::min(hstart + sizeY, inHeight + paddingH);
+            int wend = std::min(wstart + sizeX, inWidth + paddingW);
+            int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart);
+            dstart = std::max(dstart, 0);
+            hstart = std::max(hstart, 0);
+            wstart = std::max(wstart, 0);
+            dend = std::min(dend, static_cast<int>(inDepth));
+            hend = std::min(hend, static_cast<int>(inHeight));
+            wend = std::min(wend, static_cast<int>(inWidth));
+
+            CHECK(poolSize);
+            tgtData[(pd * outputH + ph) * outputW + pw] = 0;  // clear
+            for (int d = dstart; d < dend; ++d) {
+              for (int h = hstart; h < hend; ++h) {
+                for (int w = wstart; w < wend; ++w) {
+                  tgtData[(pd * outputH + ph) * outputW + pw] +=
+                      inData[(d * inHeight + h) * inWidth + w];
+                }
+              }
+            }
+            tgtData[(pd * outputH + ph) * outputW + pw] /= poolSize;
+          }
+        }
+      }
+      // compute offset
+      inData += inDepth * inHeight * inWidth;
+      tgtData += outputD * outputH * outputW;
+    }
+  }
+}
+
+void CpuMatrix::avgPool3DBackward(Matrix& input,
+                                  size_t imgSizeD,
+                                  size_t imgSizeH,
+                                  size_t imgSizeW,
+                                  size_t outputD,
+                                  size_t outputH,
+                                  size_t outputW,
+                                  size_t sizeZ,
+                                  size_t sizeY,
+                                  size_t sizeX,
+                                  size_t strideD,
+                                  size_t strideH,
+                                  size_t strideW,
+                                  size_t paddingD,
+                                  size_t paddingH,
+                                  size_t paddingW,
+                                  real scaleTargets,
+                                  real scaleOutput) {
+  size_t num = input.getHeight();
+  size_t channels = input.getWidth() / outputD / outputH / outputW;
+  CHECK(imgSizeD * imgSizeH * imgSizeW * channels == getWidth());
+  real* inData = input.getData();
+  real* outData = getData();
+
+  for (size_t n = 0; n < num; ++n) {
+    if (!input.isContiguous()) {
+      inData = input.getData() + n * input.getStride();
+    }
+    for (size_t c = 0; c < channels; ++c) {
+      for (size_t pd = 0; pd < outputD; ++pd) {
+        for (size_t ph = 0; ph < outputH; ++ph) {
+          for (size_t pw = 0; pw < outputW; ++pw) {
+            int dstart = pd * strideD - paddingD;
+            int hstart = ph * strideH - paddingH;
+            int wstart = pw * strideW - paddingW;
+            int dend = std::min(dstart + sizeZ, imgSizeD + paddingD);
+            int hend = std::min(hstart + sizeY, imgSizeH + paddingH);
+            int wend = std::min(wstart + sizeX, imgSizeW + paddingW);
+            int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart);
+            dstart = std::max(dstart, 0);
+            hstart = std::max(hstart, 0);
+            wstart = std::max(wstart, 0);
+            dend = std::min(dend, static_cast<int>(imgSizeD));
+            hend = std::min(hend, static_cast<int>(imgSizeH));
+            wend = std::min(wend, static_cast<int>(imgSizeW));
+            CHECK(poolSize);
+            for (int d = dstart; d < dend; ++d) {
+              for (int h = hstart; h < hend; ++h) {
+                for (int w = wstart; w < wend; ++w) {
+                  outData[(d * imgSizeH + h) * imgSizeW + w] +=
+                      inData[(pd * outputH + ph) * outputW + pw] / poolSize;
+                }
+              }
+            }
+          }
+        }
+      }
+      // offset
+      outData += imgSizeD * imgSizeH * imgSizeW;
+      inData += outputD * outputH * outputW;
+    }
+  }
+}
+
 /**
  * Input: one or more sequences. Each sequence contains some instances.
  * Output: output size is the number of input sequences (NOT input instances).
@@ -3975,6 +4526,95 @@ void CpuMatrix::bilinearBackward(const Matrix& out,
   }
 }
 
+void CpuMatrix::vol2Col(real* data,
+                        int channels,
+                        int depth,
+                        int height,
+                        int width,
+                        int filterD,
+                        int filterH,
+                        int filterW,
+                        int strideD,
+                        int strideH,
+                        int strideW,
+                        int paddingD,
+                        int paddingH,
+                        int paddingW) {
+  real* outData = getData();
+  int outHeight = (height + 2 * paddingH - filterH) / strideH + 1;
+  int outWidth = (width + 2 * paddingW - filterW) / strideW + 1;
+  int outDepth = (depth + 2 * paddingD - filterD) / strideD + 1;
+
+  int channelsCol = channels * filterD * filterH * filterW;
+  for (int c = 0; c < channelsCol; ++c) {
+    int wOffset = c % filterW;
+    int hOffset = (c / filterW) % filterH;
+    int dOffset = (c / filterW / filterH) % filterD;
+    int cIn = c / filterW / filterH / filterD;
+    for (int d = 0; d < outDepth; ++d) {
+      for (int h = 0; h < outHeight; ++h) {
+        for (int w = 0; w < outWidth; ++w) {
+          int dPad = d * strideD - paddingD + dOffset;
+          int hPad = h * strideH - paddingH + hOffset;
+          int wPad = w * strideW - paddingW + wOffset;
+
+          if (hPad >= 0 && hPad < height && wPad >= 0 && wPad < width &&
+              dPad >= 0 && dPad < depth)
+            outData[((c * outDepth + d) * outHeight + h) * outWidth + w] =
+                data[((cIn * depth + dPad) * height + hPad) * width + wPad];
+          else
+            outData[((c * outDepth + d) * outHeight + h) * outWidth + w] = 0;
+        }
+      }
+    }
+  }
+}
+
+void CpuMatrix::col2Vol(real* trg,
+                        int channels,
+                        int depth,
+                        int height,
+                        int width,
+                        int filterD,
+                        int filterH,
+                        int filterW,
+                        int strideD,
+                        int strideH,
+                        int strideW,
+                        int paddingD,
+                        int paddingH,
+                        int paddingW,
+                        real alpha,
+                        real beta) {
+  real* src = getData();
+  int outDepth = (depth + 2 * paddingD - filterD) / strideD + 1;
+  int outHeight = (height + 2 * paddingH - filterH) / strideH + 1;
+  int outWidth = (width + 2 * paddingW - filterW) / strideW + 1;
+  int channelsCol = channels * filterD * filterH * filterW;
+  for (int c = 0; c < channelsCol; ++c) {
+    int wOffset = c % filterW;
+    int hOffset = (c / filterW) % filterH;
+    int dOffset = (c / filterW / filterH) % filterD;
+    int cIm = c / filterW / filterH / filterD;
+    for (int d = 0; d < outDepth; ++d) {
+      for (int h = 0; h < outHeight; ++h) {
+        for (int w = 0; w < outWidth; ++w) {
+          int dPad = d * strideD - paddingD + dOffset;
+          int hPad = h * strideH - paddingH + hOffset;
+          int wPad = w * strideW - paddingW + wOffset;
+          if (hPad >= 0 && hPad < height && wPad >= 0 && wPad < width &&
+              dPad >= 0 && dPad < depth)
+            trg[((cIm * depth + dPad) * height + hPad) * width + wPad] =
+                alpha *
+                    src[((c * outDepth + d) * outHeight + h) * outWidth + w] +
+                beta *
+                    trg[((cIm * depth + dPad) * height + hPad) * width + wPad];
+        }
+      }
+    }
+  }
+}
+
 ////////////////////////////////////////////////////////////////
 //               functions executed via cpu                   //
 ////////////////////////////////////////////////////////////////

paddle/math/Matrix.h

@@ -928,15 +928,102 @@ public:
                                size_t paddingW) {
     LOG(FATAL) << "Not implemeted";
   }
-
   /**
-   * Input: one or more sequences. Each sequence contains some instances.
-   *
-   * Output: output size is the number of input sequences (NOT input
-   * instances).
-   *
-   * output[i] is set to max_input[i].
+   * Pooling 3D forward operation, pick out the largest element
+   * in the sizeX of value
    */
+  virtual void maxPool3DForward(Matrix& inputMat,
+                                Matrix& maxPoolIdx,
+                                size_t channels,
+                                size_t imgSizeD,
+                                size_t imgSizeH,
+                                size_t imgSizeW,
+                                size_t outputD,
+                                size_t outputH,
+                                size_t outputW,
+                                size_t sizeZ,
+                                size_t sizeY,
+                                size_t sizeX,
+                                size_t strideD,
+                                size_t strideH,
+                                size_t strideW,
+                                size_t paddingD,
+                                size_t paddingH,
+                                size_t paddingW) {
+    LOG(FATAL) << "Not implemeted";
+  }
+
+  virtual void maxPool3DBackward(Matrix& outGrad,
+                                 Matrix& maxPoolIdx,
+                                 size_t imgSizeD,
+                                 size_t imgSizeH,
+                                 size_t imgSizeW,
+                                 size_t outputD,
+                                 size_t outputH,
+                                 size_t outputW,
+                                 size_t sizeZ,
+                                 size_t sizeY,
+                                 size_t sizeX,
+                                 size_t strideD,
+                                 size_t strideH,
+                                 size_t strideW,
+                                 size_t paddingD,
+                                 size_t paddingH,
+                                 size_t paddingW,
+                                 real scaleTargets,
+                                 real scaleOutput) {
+    LOG(FATAL) << "Not implemeted";
+  }
+
+  virtual void avgPool3DForward(Matrix& input,
+                                size_t channels,
+                                size_t imgSizeD,
+                                size_t imgSizeH,
+                                size_t imgSizeW,
+                                size_t outputD,
+                                size_t outputH,
+                                size_t outputW,
+                                size_t sizeZ,
+                                size_t sizeY,
+                                size_t sizeX,
+                                size_t strideD,
+                                size_t strideH,
+                                size_t strideW,
+                                size_t paddingD,
+                                size_t paddingH,
+                                size_t paddingW) {
+    LOG(FATAL) << "Not implemeted";
+  }
+
+  virtual void avgPool3DBackward(Matrix& input,
+                                 size_t imgSizeD,
+                                 size_t imgSizeH,
+                                 size_t imgSizeW,
+                                 size_t outputD,
+                                 size_t outputH,
+                                 size_t outputW,
+                                 size_t sizeZ,
+                                 size_t sizeY,
+                                 size_t sizeX,
+                                 size_t strideD,
+                                 size_t strideH,
+                                 size_t strideW,
+                                 size_t paddingD,
+                                 size_t paddingH,
+                                 size_t paddingW,
+                                 real scaleTargets,
+                                 real scaleOutput) {
+    LOG(FATAL) << "Not implemeted";
+  }
+
+  /**
+ * Input: one or more sequences. Each sequence contains some instances.
+ *
+ * Output: output size is the number of input sequences (NOT input
+ * instances).
+ *
+ * output[i] is set to max_input[i].
+ */
   virtual void maxSequenceForward(Matrix& input,
                                   const IVector& sequence,
                                   IVector& index) {
@@ -1039,6 +1126,42 @@ public:
     LOG(FATAL) << "Not implemented";
   }
 
+  virtual void vol2Col(real* data,
+                       int channels,
+                       int depth,
+                       int height,
+                       int width,
+                       int filterD,
+                       int filterH,
+                       int filterW,
+                       int strideD,
+                       int strideH,
+                       int strideW,
+                       int paddingD,
+                       int paddingH,
+                       int paddingW) {
+    LOG(FATAL) << "Not implemeted";
+  }
+
+  virtual void col2Vol(real* trg,
+                       int channels,
+                       int depth,
+                       int height,
+                       int width,
+                       int filterD,
+                       int filterH,
+                       int filterW,
+                       int strideD,
+                       int strideH,
+                       int strideW,
+                       int paddingD,
+                       int paddingH,
+                       int paddingW,
+                       real alpha,
+                       real beta) {
+    LOG(FATAL) << "Not implemeted";
+  }
+
   virtual void bilinearForward(const Matrix& in,
                                const size_t inImgH,
                                const size_t inImgW,
@@ -1348,6 +1471,82 @@ public:
                        size_t paddingH,
                        size_t paddingW);
 
+  void maxPool3DForward(Matrix& inputMat,
+                        Matrix& maxPoolIdx,
+                        size_t channels,
+                        size_t imgSizeD,
+                        size_t imgSizeH,
+                        size_t imgSizeW,
+                        size_t outputD,
+                        size_t outputH,
+                        size_t outputW,
+                        size_t sizeZ,
+                        size_t sizeY,
+                        size_t sizeX,
+                        size_t strideD,
+                        size_t strideH,
+                        size_t strideW,
+                        size_t paddingD,
+                        size_t paddingH,
+                        size_t paddingW);
+
+  void maxPool3DBackward(Matrix& outGrad,
+                         Matrix& maxPoolIdx,
+                         size_t imgSizeD,
+                         size_t imgSizeH,
+                         size_t imgSizeW,
+                         size_t outputD,
+                         size_t outputH,
+                         size_t outputW,
+                         size_t sizeZ,
+                         size_t sizeY,
+                         size_t sizeX,
+                         size_t strideD,
+                         size_t strideH,
+                         size_t strideW,
+                         size_t paddingD,
+                         size_t paddingH,
+                         size_t paddingW,
+                         real scaleTargets,
+                         real scaleOutput);
+
+  void avgPool3DForward(Matrix& input,
+                        size_t channels,
+                        size_t imgSizeD,
+                        size_t imgSizeH,
+                        size_t imgSizeW,
+                        size_t outputD,
+                        size_t outputH,
+                        size_t outputW,
+                        size_t sizeZ,
+                        size_t sizeY,
+                        size_t sizeX,
+                        size_t strideD,
+                        size_t strideH,
+                        size_t strideW,
+                        size_t paddingD,
+                        size_t paddingH,
+                        size_t paddingW);
+
+  void avgPool3DBackward(Matrix& input,
+                         size_t imgSizeD,
+                         size_t imgSizeH,
+                         size_t imgSizeW,
+                         size_t outputD,
+                         size_t outputH,
+                         size_t outputW,
+                         size_t sizeZ,
+                         size_t sizeY,
+                         size_t sizeX,
+                         size_t strideD,
+                         size_t strideH,
+                         size_t strideW,
+                         size_t paddingD,
+                         size_t paddingH,
+                         size_t paddingW,
+                         real scaleTargets,
+                         real scaleOutput);
+
   void maxSequenceForward(Matrix& input,
                           const IVector& sequence,
                           IVector& index);
@@ -1374,6 +1573,38 @@ public:
                         const real ratioH,
                         const real ratioW);
 
+  void vol2Col(real* data,
+               int channels,
+               int depth,
+               int height,
+               int width,
+               int filterD,
+               int filterH,
+               int filterW,
+               int strideD,
+               int strideH,
+               int strideW,
+               int paddingD,
+               int paddingH,
+               int paddingW);
+
+  void col2Vol(real* trg,
+               int channels,
+               int depth,
+               int height,
+               int width,
+               int filterD,
+               int filterH,
+               int filterW,
+               int strideD,
+               int strideH,
+               int strideW,
+               int paddingD,
+               int paddingH,
+               int paddingW,
+               real alpha,
+               real beta);
+
   void multiBinaryLabelCrossEntropy(Matrix& output, Matrix& label);
 
   void multiBinaryLabelCrossEntropyBp(Matrix& output, Matrix& label);
@@ -1507,6 +1738,82 @@ public:
                        size_t paddingH,
                        size_t paddingW);
 
+  void maxPool3DForward(Matrix& inputMat,
+                        Matrix& maxPoolIdx,
+                        size_t channels,
+                        size_t imgSizeD,
+                        size_t imgSizeH,
+                        size_t imgSizeW,
+                        size_t outputD,
+                        size_t outputH,
+                        size_t outputW,
+                        size_t sizeZ,
+                        size_t sizeY,
+                        size_t sizeX,
+                        size_t strideD,
+                        size_t strideH,
+                        size_t strideW,
+                        size_t paddingD,
+                        size_t paddingH,
+                        size_t paddingW);
+
+  void maxPool3DBackward(Matrix& outGrad,
+                         Matrix& maxPoolIdx,
+                         size_t imgSizeD,
+                         size_t imgSizeH,
+                         size_t imgSizeW,
+                         size_t outputD,
+                         size_t outputH,
+                         size_t outputW,
+                         size_t sizeZ,
+                         size_t sizeY,
+                         size_t sizeX,
+                         size_t strideD,
+                         size_t strideH,
+                         size_t strideW,
+                         size_t paddingD,
+                         size_t paddingH,
+                         size_t paddingW,
+                         real scaleTargets,
+                         real scaleOutput);
+
+  void avgPool3DForward(Matrix& input,
+                        size_t channels,
+                        size_t imgSizeD,
+                        size_t imgSizeH,
+                        size_t imgSizeW,
+                        size_t outputD,
+                        size_t outputH,
+                        size_t outputW,
+                        size_t sizeZ,
+                        size_t sizeY,
+                        size_t sizeX,
+                        size_t strideD,
+                        size_t strideH,
+                        size_t strideW,
+                        size_t paddingD,
+                        size_t paddingH,
+                        size_t paddingW);
+
+  void avgPool3DBackward(Matrix& input,
+                         size_t imgSizeD,
+                         size_t imgSizeH,
+                         size_t imgSizeW,
+                         size_t outputD,
+                         size_t outputH,
+                         size_t outputW,
+                         size_t sizeZ,
+                         size_t sizeY,
+                         size_t sizeX,
+                         size_t strideD,
+                         size_t strideH,
+                         size_t strideW,
+                         size_t paddingD,
+                         size_t paddingH,
+                         size_t paddingW,
+                         real scaleTargets,
+                         real scaleOutput);
+
   void maxSequenceForward(Matrix& input,
                           const IVector& sequence,
                           IVector& index);
@@ -1715,6 +2022,38 @@ public:
                         const real ratioH,
                         const real ratioW);
 
+  void vol2Col(real* data,
+               int channels,
+               int depth,
+               int height,
+               int width,
+               int filterD,
+               int filterH,
+               int filterW,
+               int strideD,
+               int strideH,
+               int strideW,
+               int paddingD,
+               int paddingH,
+               int paddingW);
+
+  void col2Vol(real* trg,
+               int channels,
+               int depth,
+               int height,
+               int width,
+               int filterD,
+               int filterH,
+               int filterW,
+               int strideD,
+               int strideH,
+               int strideW,
+               int paddingD,
+               int paddingH,
+               int paddingW,
+               real alpha,
+               real beta);
+
   template <typename ExpressionType>
   void operator=(const ExpressionType& expr) {
     TensorCpuApply<real>(*this, expr);

paddle/math/tests/test_matrixCompare.cpp

@@ -18,6 +18,7 @@ limitations under the License. */
 
 #include <gtest/gtest.h>
 #include "TensorCheck.h"
+#include "paddle/math/MathUtils.h"
 #include "paddle/math/Matrix.h"
 #include "paddle/math/SparseMatrix.h"
 #include "paddle/testing/TestUtil.h"
@@ -1203,4 +1204,497 @@ TEST(Matrix, warpCTC) {
   }
 }
 
+void testMaxPool3DFwdBwd(int numSamples,
+                         int channels,
+                         int imgSizeD,
+                         int imgSizeH,
+                         int imgSizeW,
+                         int ksizeD,
+                         int ksizeH,
+                         int ksizeW,
+                         int strideD,
+                         int strideH,
+                         int strideW,
+                         int padD,
+                         int padH,
+                         int padW) {
+  int outD = outputSize(imgSizeD, ksizeD, padD, strideD, true);
+  int outH = outputSize(imgSizeH, ksizeH, padH, strideH, true);
+  int outW = outputSize(imgSizeW, ksizeW, padW, strideW, true);
+
+  int inWidth = channels * imgSizeD * imgSizeH * imgSizeW;
+  MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false);
+  MatrixPtr inputGpu = GpuMatrix::create(numSamples, inWidth, false, true);
+
+  int outWidth = channels * outD * outH * outW;
+  MatrixPtr target = CpuMatrix::create(numSamples, outWidth, false, false);
+  MatrixPtr targetGpu = GpuMatrix::create(numSamples, outWidth, false, true);
+  MatrixPtr maxIdx = CpuMatrix::create(numSamples, outWidth, false, false);
+  MatrixPtr maxIdxGpu = GpuMatrix::create(numSamples, outWidth, false, true);
+
+  input->randomizeUniform();
+  target->randomizeUniform();
+  inputGpu->copyFrom(*input);
+  targetGpu->copyFrom(*target);
+
+  target->maxPool3DForward(*input,
+                           *maxIdx,
+                           channels,
+                           imgSizeD,
+                           imgSizeH,
+                           imgSizeW,
+                           outD,
+                           outH,
+                           outW,
+                           ksizeD,
+                           ksizeH,
+                           ksizeW,
+                           strideD,
+                           strideH,
+                           strideW,
+                           padD,
+                           padH,
+                           padW);
+  targetGpu->maxPool3DForward(*inputGpu,
+                              *maxIdxGpu,
+                              channels,
+                              imgSizeD,
+                              imgSizeH,
+                              imgSizeW,
+                              outD,
+                              outH,
+                              outW,
+                              ksizeD,
+                              ksizeH,
+                              ksizeW,
+                              strideD,
+                              strideH,
+                              strideW,
+                              padD,
+                              padH,
+                              padW);
+  MatrixPtr targetCheck = CpuMatrix::create(numSamples, outWidth, false, false);
+  targetCheck->copyFrom(*targetGpu);
+  checkMatrixEqual(target, targetCheck);
+
+  MatrixPtr inputGrad = CpuMatrix::create(numSamples, inWidth, false, false);
+  MatrixPtr inputGpuGrad = GpuMatrix::create(numSamples, inWidth, false, true);
+  MatrixPtr targetGrad = CpuMatrix::create(numSamples, outWidth, false, false);
+  MatrixPtr targetGpuGrad =
+      GpuMatrix::create(numSamples, outWidth, false, true);
+
+  inputGrad->randomizeUniform();
+  targetGrad->randomizeUniform();
+  inputGpuGrad->copyFrom(*inputGrad);
+  targetGpuGrad->copyFrom(*targetGrad);
+
+  inputGrad->maxPool3DBackward(*targetGrad,
+                               *maxIdx,
+                               imgSizeD,
+                               imgSizeH,
+                               imgSizeW,
+                               outD,
+                               outH,
+                               outW,
+                               ksizeD,
+                               ksizeH,
+                               ksizeW,
+                               strideD,
+                               strideH,
+                               strideW,
+                               padD,
+                               padH,
+                               padW,
+                               1.0,
+                               1.0);
+  inputGpuGrad->maxPool3DBackward(*targetGpuGrad,
+                                  *maxIdxGpu,
+                                  imgSizeD,
+                                  imgSizeH,
+                                  imgSizeW,
+                                  outD,
+                                  outH,
+                                  outW,
+                                  ksizeD,
+                                  ksizeH,
+                                  ksizeW,
+                                  strideD,
+                                  strideH,
+                                  strideW,
+                                  padD,
+                                  padH,
+                                  padW,
+                                  1.0,
+                                  1.0);
+  MatrixPtr targetBwdCheck =
+      CpuMatrix::create(numSamples, inWidth, false, false);
+  targetBwdCheck->copyFrom(*inputGpuGrad);
+  checkMatrixEqual(inputGrad, targetBwdCheck);
+}
+
+void testAvgPool3DFwdBwd(int numSamples,
+                         int channels,
+                         int imgSizeD,
+                         int imgSizeH,
+                         int imgSizeW,
+                         int ksizeD,
+                         int ksizeH,
+                         int ksizeW,
+                         int strideD,
+                         int strideH,
+                         int strideW,
+                         int padD,
+                         int padH,
+                         int padW) {
+  int outD = outputSize(imgSizeD, ksizeD, padD, strideD, true);
+  int outH = outputSize(imgSizeH, ksizeH, padH, strideH, true);
+  int outW = outputSize(imgSizeW, ksizeW, padW, strideW, true);
+
+  int inWidth = imgSizeD * imgSizeH * imgSizeW * channels;
+  MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false);
+  MatrixPtr inputGpu = GpuMatrix::create(numSamples, inWidth, false, true);
+
+  int outWidth = channels * outD * outH * outW;
+  MatrixPtr target = CpuMatrix::create(numSamples, outWidth, false, false);
+  MatrixPtr targetGpu = GpuMatrix::create(numSamples, outWidth, false, true);
+
+  input->randomizeUniform();
+  target->randomizeUniform();
+  inputGpu->copyFrom(*input);
+  targetGpu->copyFrom(*target);
+
+  target->avgPool3DForward(*input,
+                           channels,
+                           imgSizeD,
+                           imgSizeH,
+                           imgSizeW,
+                           outD,
+                           outH,
+                           outW,
+                           ksizeD,
+                           ksizeH,
+                           ksizeW,
+                           strideD,
+                           strideH,
+                           strideW,
+                           padD,
+                           padH,
+                           padW);
+
+  targetGpu->avgPool3DForward(*inputGpu,
+                              channels,
+                              imgSizeD,
+                              imgSizeH,
+                              imgSizeW,
+                              outD,
+                              outH,
+                              outW,
+                              ksizeD,
+                              ksizeH,
+                              ksizeW,
+                              strideD,
+                              strideH,
+                              strideW,
+                              padD,
+                              padH,
+                              padW);
+
+  TensorCheckErr(*target, *targetGpu);
+
+  MatrixPtr inputGrad = CpuMatrix::create(numSamples, inWidth, false, false);
+  MatrixPtr inputGpuGrad = GpuMatrix::create(numSamples, inWidth, false, true);
+  MatrixPtr targetGrad = CpuMatrix::create(numSamples, outWidth, false, false);
+  MatrixPtr targetGpuGrad =
+      GpuMatrix::create(numSamples, outWidth, false, true);
+
+  inputGrad->randomizeUniform();
+  targetGrad->randomizeUniform();
+  inputGpuGrad->copyFrom(*inputGrad);
+  targetGpuGrad->copyFrom(*targetGrad);
+
+  inputGrad->avgPool3DBackward(*targetGrad,
+                               imgSizeD,
+                               imgSizeH,
+                               imgSizeW,
+                               outD,
+                               outH,
+                               outW,
+                               ksizeD,
+                               ksizeH,
+                               ksizeW,
+                               strideD,
+                               strideH,
+                               strideW,
+                               padD,
+                               padH,
+                               padW,
+                               1.0,
+                               1.0);
+
+  inputGpuGrad->avgPool3DBackward(*targetGpuGrad,
+                                  imgSizeD,
+                                  imgSizeH,
+                                  imgSizeW,
+                                  outD,
+                                  outH,
+                                  outW,
+                                  ksizeD,
+                                  ksizeH,
+                                  ksizeW,
+                                  strideD,
+                                  strideH,
+                                  strideW,
+                                  padD,
+                                  padH,
+                                  padW,
+                                  1.0,
+                                  1.0);
+  TensorCheckErr(*inputGrad, *inputGpuGrad);
+}
+
+// TODO(yi): I noticed many such blindly combinatorial tests in this
+// file.  They are no help to locate defects at all.
+TEST(Matrix, Pool3DFwdBwd) {
+  for (auto numSamples : {1, 3}) {
+    for (auto channels : {3}) {
+      for (auto imgSizeD : {9, 16}) {
+        for (auto imgSizeH : {9, 32}) {
+          for (auto imgSizeW : {9, 32}) {
+            for (auto sizeX : {3}) {
+              for (auto sizeY : {3}) {
+                for (auto sizeZ : {3}) {
+                  for (auto sD : {2}) {
+                    for (auto sH : {2}) {
+                      for (auto sW : {2}) {
+                        for (auto pD : {0, (sizeZ - 1) / 2}) {
+                          for (auto pH : {0, (sizeY - 1) / 2}) {
+                            for (auto pW : {0, (sizeX - 1) / 2}) {
+                              VLOG(3) << " numSamples=" << numSamples
+                                      << " channels=" << channels
+                                      << " imgSizeD=" << imgSizeD
+                                      << " imgSizeH=" << imgSizeH
+                                      << " imgSizeW=" << imgSizeW
+                                      << " sizeX=" << sizeX
+                                      << " sizeY=" << sizeY
+                                      << " sizeZ=" << sizeZ << " strideD=" << sD
+                                      << " strideH=" << sH << " strideW=" << sW
+                                      << " padingD=" << pD << " padingH=" << pH
+                                      << " padingW=" << pW;
+
+                              testMaxPool3DFwdBwd(numSamples,
+                                                  channels,
+                                                  imgSizeD,
+                                                  imgSizeH,
+                                                  imgSizeW,
+                                                  sizeX,
+                                                  sizeY,
+                                                  sizeZ,
+                                                  sD,
+                                                  sH,
+                                                  sW,
+                                                  pD,
+                                                  pH,
+                                                  pW);
+                              testAvgPool3DFwdBwd(numSamples,
+                                                  channels,
+                                                  imgSizeD,
+                                                  imgSizeH,
+                                                  imgSizeW,
+                                                  sizeX,
+                                                  sizeY,
+                                                  sizeZ,
+                                                  sD,
+                                                  sH,
+                                                  sW,
+                                                  pD,
+                                                  pH,
+                                                  pW);
+                            }
+                          }
+                        }
+                      }
+                    }
+                  }
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+
+  //  for (auto numSamples : {1, 3}) {
+  //    for (auto channels : {1, 3}) {
+  //      for (auto imgSizeD : {9,16}) {
+  //      for (auto imgSizeH : {9, 32}) {
+  //        for (auto imgSizeW : {9, 32}) {
+  //          for (auto sizeX : {2, 3}) {
+  //            for (auto sizeY : {2, 3}) {
+  //            for (auto sizeZ : {2,3}){
+  //              for (auto sD : {1, 2}) {
+  //              for (auto sH : {1, 2}) {
+  //                for (auto sW : {1, 2}) {
+  //                  for (auto pD : {0, (sizeZ - 1) / 2}){
+  //                  for (auto pH : {0, (sizeY - 1) / 2}) {
+  //                    for (auto pW : {0, (sizeX - 1) / 2}) {
+  //                      VLOG(3) << " numSamples=" << numSamples
+  //                              << " channels=" << channels
+  //                              << " imgSizeD=" << imgSizeD
+  //                              << " imgSizeH=" << imgSizeH
+  //                              << " imgSizeW=" << imgSizeW
+  //                              << " sizeX=" << sizeX
+  //                              << " sizeY=" << sizeY
+  //                              << " sizeZ=" << sizeZ
+  //                              << " strideD=" << sD
+  //                              << " strideH=" << sH
+  //                              << " strideW=" << sW
+  //                              << " padingD=" << pD
+  //                              << " padingH=" << pH
+  //                              << " padingW=" << pW;
+  //
+  //                      testMaxPool3DFwdBwd(numSamples,
+  //                                        channels,
+  //                                        imgSizeD,
+  //                                        imgSizeH,
+  //                                        imgSizeW,
+  //                                        sizeX,
+  //                                        sizeY,
+  //                                        sizeZ,
+  //                                        sD,
+  //                                        sH,
+  //                                        sW,
+  //                                        pD,
+  //                                        pH,
+  //                                        pW);
+  //                      testAvgPool3DFwdBwd(numSamples,
+  //                                        channels,
+  //                                        imgSizeD,
+  //                                        imgSizeH,
+  //                                        imgSizeW,
+  //                                        sizeX,
+  //                                        sizeY,
+  //                                        sizeZ,
+  //                                        sD,
+  //                                        sH,
+  //                                        sW,
+  //                                        pD,
+  //                                        pH,
+  //                                        pW);
+  //                    }
+  //                  }
+  //                }
+  //              }
+  //            }
+  //            }
+  //          }
+  //        }
+  //      }
+  //      }
+  //    }
+  //    }
+  //  }
+  //  }
+}
+
+void testMatrixCol2Vol(int depth, int height, int width) {
+  int channel = 3;
+  int filterX = 3, filterY = 4, filterZ = 5;
+  int strideX = 2, strideY = 2, strideZ = 2;
+  int padX = 1, padY = 1, padZ = 1;
+
+  MatrixPtr cpuImage =
+      std::make_shared<CpuMatrix>(channel, depth * height * width);
+  MatrixPtr gpuImage =
+      std::make_shared<GpuMatrix>(channel, depth * height * width);
+  cpuImage->randomizeUniform();
+  gpuImage->copyFrom(*cpuImage);
+
+  int outD = outputSize(depth, filterZ, padZ, strideZ, true);
+  int outH = outputSize(height, filterY, padY, strideY, true);
+  int outW = outputSize(width, filterX, padX, strideX, true);
+
+  int colBufHeight = channel * filterZ * filterY * filterX;
+  int colBufWidth = outD * outH * outW;
+  MatrixPtr cpuColBuf = std::make_shared<CpuMatrix>(colBufHeight, colBufWidth);
+  MatrixPtr gpuColBuf = std::make_shared<GpuMatrix>(colBufHeight, colBufWidth);
+  cpuColBuf->vol2Col(cpuImage->getData(),
+                     channel,
+                     depth,
+                     height,
+                     width,
+                     filterZ,
+                     filterY,
+                     filterX,
+                     strideZ,
+                     strideY,
+                     strideX,
+                     padZ,
+                     padY,
+                     padX);
+  gpuColBuf->vol2Col(gpuImage->getData(),
+                     channel,
+                     depth,
+                     height,
+                     width,
+                     filterZ,
+                     filterY,
+                     filterX,
+                     strideZ,
+                     strideY,
+                     strideX,
+                     padZ,
+                     padY,
+                     padX);
+  TensorCheckEqual(*cpuColBuf, *gpuColBuf);
+
+  cpuColBuf->randomizeUniform();
+  gpuColBuf->copyFrom(*cpuColBuf);
+  cpuColBuf->col2Vol(cpuImage->getData(),
+                     channel,
+                     depth,
+                     height,
+                     width,
+                     filterZ,
+                     filterY,
+                     filterX,
+                     strideZ,
+                     strideY,
+                     strideX,
+                     padZ,
+                     padY,
+                     padX,
+                     1.0,
+                     1.0);
+  gpuColBuf->col2Vol(gpuImage->getData(),
+                     channel,
+                     depth,
+                     height,
+                     width,
+                     filterZ,
+                     filterY,
+                     filterX,
+                     strideZ,
+                     strideY,
+                     strideX,
+                     padZ,
+                     padY,
+                     padX,
+                     1.0,
+                     1.0);
+  TensorCheckErr(*cpuImage, *gpuImage);
+}
+
+TEST(Matrix, col2Vol) {
+  for (auto depth : {9, 16, 64}) {
+    for (auto height : {9, 11, 128}) {
+      for (auto width : {9, 32, 128}) {
+        VLOG(3) << "depth=" << depth << " height=" << height
+                << " width=" << width;
+        testMatrixCol2Vol(depth, height, width);
+      }
+    }
+  }
+}
+
 #endif

paddle/operators/CMakeLists.txt

+file(GLOB GENERAL_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*_op.cc")
+string(REPLACE ".cc" "" GENERAL_OPS "${GENERAL_OPS}")
 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
     # for ops.
+    set(OP_LIBRARY ${TARGET} ${OP_LIBRARY} PARENT_SCOPE)
     set(cc_srcs)
     set(cu_srcs)
     set(op_common_deps operator op_registry)
@@ -43,33 +46,26 @@ endfunction()
 
 add_subdirectory(math)
 
-cc_test(gather_test SRCS gather_test.cc DEPS tensor)
-op_library(gather_op SRCS gather_op.cc gather_op.cu)
-
-cc_test(scatter_test SRCS scatter_test.cc DEPS tensor)
-op_library(scatter_op SRCS scatter_op.cc scatter_op.cu)
-
-cc_library(net_op SRCS net_op.cc DEPS op_registry)
-cc_test(net_op_test SRCS net_op_test.cc DEPS net_op)
-
-op_library(add_op SRCS add_op.cc add_op.cu)
-
-op_library(mean_op SRCS mean_op.cc mean_op.cu)
+list(REMOVE_ITEM GENERAL_OPS
+     net_op
+     minus_op
+     mul_op
+     recurrent_op
+     scale_op)
 
+op_library(net_op SRCS net_op.cc)
+op_library(minus_op SRCS minus_op.cc minus_op.cu DEPS scale_op)
 op_library(mul_op SRCS mul_op.cc mul_op.cu DEPS math_function)
-op_library(rowwise_add_op SRCS rowwise_add_op.cu rowwise_add_op.cc)
+op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc 
+  DEPS framework_proto tensor operator net_op)
+op_library(scale_op SRCS scale_op.cc scale_op.cu DEPS net_op)
 
-op_library(sigmoid_op SRCS sigmoid_op.cc sigmoid_op.cu)
-op_library(softmax_op SRCS softmax_op.cc softmax_op.cu)
-op_library(gaussian_random_op SRCS gaussian_random_op.cc gaussian_random_op.cu)
-op_library(cross_entropy_op SRCS cross_entropy_op.cc cross_entropy_op.cu)
-op_library(fill_zeros_like_op SRCS fill_zeros_like_op.cc fill_zeros_like_op.cu)
+foreach(src ${GENERAL_OPS})
+    op_library(${src} SRCS ${src}.cc ${src}.cu)
+endforeach()
 
-op_library(sgd_op SRCS sgd_op.cc sgd_op.cu)
+set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library")
 
-op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc
-    DEPS framework_proto tensor op_registry operator net_op)
-op_library(uniform_random_op SRCS uniform_random_op.cc uniform_random_op.cu)
-op_library(lookup_table_op SRCS lookup_table_op.cc lookup_table_op.cu)
-op_library(scale_op SRCS scale_op.cc scale_op.cu DEPS net_op)
-op_library(minus_op SRCS minus_op.cc minus_op.cu DEPS scale_op)
+cc_test(gather_test SRCS gather_test.cc DEPS tensor)
+cc_test(net_op_test SRCS net_op_test.cc DEPS net_op)
+cc_test(scatter_test SRCS scatter_test.cc DEPS tensor)

paddle/operators/add_op.cc

@@ -57,7 +57,7 @@ class AddOpGrad : public framework::OperatorWithKernel {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP(add_two, ops::AddOp, ops::AddOpMaker, add_two_grad, ops::AddOpGrad);
+REGISTER_OP(add_two, ops::AddOp, ops::AddOpMaker, ops::AddOpGrad);
 
 REGISTER_OP_CPU_KERNEL(add_two,
                        ops::AddKernel<paddle::platform::CPUPlace, float>);

paddle/operators/cross_entropy_op.cc

@@ -67,8 +67,7 @@ OnehotCrossEntropy Operator.
 
 namespace ops = paddle::operators;
 REGISTER_OP(onehot_cross_entropy, ops::OnehotCrossEntropyOp,
-            ops::OnehotCrossEntropyOpMaker, onehot_cross_entropy_grad,
-            ops::OnehotCrossEntropyGradientOp);
+            ops::OnehotCrossEntropyOpMaker, ops::OnehotCrossEntropyGradientOp);
 REGISTER_OP_CPU_KERNEL(onehot_cross_entropy,
                        ops::OnehotCrossEntropyOpKernel<float>);
 REGISTER_OP_CPU_KERNEL(onehot_cross_entropy_grad,

paddle/operators/gather_op.cc

@@ -63,8 +63,7 @@ Out = X[Index]
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP(gather, ops::GatherOp, ops::GatherOpMaker, gather_grad,
-            ops::GatherGradOp);
+REGISTER_OP(gather, ops::GatherOp, ops::GatherOpMaker, ops::GatherGradOp);
 REGISTER_OP_CPU_KERNEL(gather,
                        ops::GatherOpKernel<paddle::platform::CPUPlace, float>);
 REGISTER_OP_CPU_KERNEL(

paddle/operators/lookup_table_op.cc

@@ -66,7 +66,7 @@ class LookupTableOpGrad : public framework::OperatorWithKernel {
 
 namespace ops = paddle::operators;
 REGISTER_OP(lookup_table, ops::LookupTableOp, ops::LookupTableOpMaker,
-            lookup_table_grad, ops::LookupTableOpGrad);
+            ops::LookupTableOpGrad);
 
 REGISTER_OP_CPU_KERNEL(lookup_table, ops::LookupTableKernel<float>);
 REGISTER_OP_CPU_KERNEL(lookup_table_grad, ops::LookupTableGradKernel<float>);

paddle/operators/mean_op.cc

@@ -54,7 +54,7 @@ class MeanGradOp : public framework::OperatorWithKernel {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP(mean, ops::MeanOp, ops::MeanOpMaker, mean_grad, ops::MeanGradOp);
+REGISTER_OP(mean, ops::MeanOp, ops::MeanOpMaker, ops::MeanGradOp);
 REGISTER_OP_CPU_KERNEL(mean,
                        ops::MeanKernel<paddle::platform::CPUPlace, float>);
 REGISTER_OP_CPU_KERNEL(mean_grad,

paddle/operators/minus_op.cc

@@ -81,7 +81,6 @@ class MinusGradOp : public NetOp {
 USE_OP(scale);
 USE_OP_ITSELF(identity);
 namespace ops = paddle::operators;
-REGISTER_OP(minus, ops::MinusOp, ops::MinusOpMaker, minus_grad,
-            ops::MinusGradOp<float>);
+REGISTER_OP(minus, ops::MinusOp, ops::MinusOpMaker, ops::MinusGradOp<float>);
 REGISTER_OP_CPU_KERNEL(minus,
                        ops::MinusKernel<paddle::platform::CPUPlace, float>);

paddle/operators/mul_op.cc

@@ -84,7 +84,7 @@ class MulOpGrad : public framework::OperatorWithKernel {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad);
+REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, ops::MulOpGrad);
 REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUPlace, float>);
 REGISTER_OP_CPU_KERNEL(mul_grad,
                        ops::MulGradKernel<paddle::platform::CPUPlace, float>);

paddle/operators/rowwise_add_op.cc

@@ -74,7 +74,7 @@ class RowwiseAddGradOp : public framework::OperatorWithKernel {
 
 namespace ops = paddle::operators;
 REGISTER_OP(rowwise_add, ops::RowwiseAddOp, ops::RowwiseAddOpMaker,
-            rowwise_add_grad, ops::RowwiseAddGradOp);
+            ops::RowwiseAddGradOp);
 REGISTER_OP_CPU_KERNEL(
     rowwise_add, ops::RowwiseAddKernel<paddle::platform::CPUPlace, float>);
 REGISTER_OP_CPU_KERNEL(

paddle/operators/scale_op.cc

@@ -97,7 +97,7 @@ class IdentityOp : public NetOp {
 
 namespace ops = paddle::operators;
 
-REGISTER_OP(scale, ops::ScaleOp, ops::ScaleOpMaker<float>, scale_grad,
+REGISTER_OP(scale, ops::ScaleOp, ops::ScaleOpMaker<float>,
             ops::ScaleGradOp<float>);
 REGISTER_OP_CPU_KERNEL(scale,
                        ops::ScaleKernel<paddle::platform::CPUPlace, float>);

paddle/operators/scatter_op.cc

@@ -77,8 +77,7 @@ Out[Index] = Ref[Index] + Updates
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP(scatter, ops::ScatterOp, ops::ScatterOpMaker, scatter_grad,
-            ops::ScatterGradOp);
+REGISTER_OP(scatter, ops::ScatterOp, ops::ScatterOpMaker, ops::ScatterGradOp);
 REGISTER_OP_CPU_KERNEL(scatter,
                        ops::ScatterOpKernel<paddle::platform::CPUPlace, float>);
 REGISTER_OP_CPU_KERNEL(

paddle/operators/sigmoid_op.cc

@@ -53,8 +53,7 @@ class SigmoidOpGrad : public framework::OperatorWithKernel {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP(sigmoid, ops::SigmoidOp, ops::SigmoidOpMaker, sigmoid_grad,
-            ops::SigmoidOpGrad);
+REGISTER_OP(sigmoid, ops::SigmoidOp, ops::SigmoidOpMaker, ops::SigmoidOpGrad);
 REGISTER_OP_CPU_KERNEL(sigmoid,
                        ops::SigmoidKernel<paddle::platform::CPUPlace, float>);
 REGISTER_OP_CPU_KERNEL(

paddle/operators/softmax_op.cc

@@ -62,8 +62,7 @@ class SoftmaxOpGrad : public framework::OperatorWithKernel {
 
 namespace ops = paddle::operators;
 
-REGISTER_OP(softmax, ops::SoftmaxOp, ops::SoftmaxOpMaker, softmax_grad,
-            ops::SoftmaxOpGrad);
+REGISTER_OP(softmax, ops::SoftmaxOp, ops::SoftmaxOpMaker, ops::SoftmaxOpGrad);
 REGISTER_OP_CPU_KERNEL(softmax,
                        ops::SoftmaxKernel<paddle::platform::CPUPlace, float>);
 REGISTER_OP_CPU_KERNEL(

paddle/parameter/Argument.cpp

@@ -186,6 +186,7 @@ void Argument::resizeAndCopyFrom(const Argument& src,
   resizeAndCopy(strs, src.strs, useGpu, stream);
   frameWidth = src.frameWidth;
   frameHeight = src.frameHeight;
+  frameDepth = src.frameDepth;
 }
 
 int32_t Argument::resizeAndCopyFrom(const Argument& src,
@@ -206,6 +207,7 @@ int32_t Argument::resizeAndCopyFrom(const Argument& src,
   dataId = src.dataId;
   frameWidth = src.frameWidth;
   frameHeight = src.frameHeight;
+  frameDepth = src.frameDepth;
 
   if (!src.sequenceStartPositions) {
     // non-sequence input, copy samples directly
@@ -677,6 +679,7 @@ void Argument::reorganizeSeqInfo(
     const ICpuGpuVectorPtr subSeqStartPos,
     std::vector<std::vector<int>>& reorganizedSeqInfo) {
   CHECK(seqStartPos);
+  reorganizedSeqInfo.clear();
 
   int seqNum = seqStartPos->getSize() - 1;
   int* seqStarts = seqStartPos->getMutableData(false);

paddle/parameter/Argument.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.
@@ -35,6 +32,7 @@ struct Argument {
         strs(nullptr),
         frameHeight(0),
         frameWidth(0),
+        frameDepth(0),
         sequenceStartPositions(nullptr),
         subSequenceStartPositions(nullptr),
         cpuSequenceDims(nullptr),
@@ -64,6 +62,7 @@ struct Argument {
     allCount = argument.allCount;
     frameHeight = argument.frameHeight;
     frameWidth = argument.frameWidth;
+    frameDepth = argument.frameDepth;
     dataId = argument.dataId;
   }
 
@@ -76,6 +75,7 @@ struct Argument {
   // A dataBatch includes batchSize frames, one frame maybe not only vector
   size_t frameHeight;
   size_t frameWidth;
+  size_t frameDepth;
 
   // If NULL, each position is treated independently.
   // Otherwise, its size should be #NumberOfSequences + 1.
@@ -136,8 +136,10 @@ struct Argument {
   }
   size_t getFrameHeight() const { return frameHeight; }
   size_t getFrameWidth() const { return frameWidth; }
+  size_t getFrameDepth() const { return frameDepth; }
   void setFrameHeight(size_t h) { frameHeight = h; }
   void setFrameWidth(size_t w) { frameWidth = w; }
+  void setFrameDepth(size_t d) { frameDepth = d; }
 
   int64_t getNumSequences() const {
     return sequenceStartPositions ? sequenceStartPositions->getSize() - 1

paddle/parameter/Parameter.h

@@ -281,7 +281,11 @@ public:
   /**
    * @brief Set the format in header.
    */
-  void setHeaderFormat(int32_t fmt) { headerFormat_ = fmt; }
+  void setHeaderFormat(int32_t fmt) {
+    CHECK(isHeaderFormatSupported(fmt)) << "Unsupported format version: "
+                                        << fmt;
+    headerFormat_ = fmt;
+  }
 
   /**
    * @brief  Parameter Update Hook.

paddle/pserver/LightNetwork.cpp

@@ -22,7 +22,6 @@ limitations under the License. */
 
 #include <arpa/inet.h>
 #include <net/if.h>
-#include <net/if_arp.h>
 #include <sys/ioctl.h>
 #include <sstream>
 

paddle/pybind/CMakeLists.txt

@@ -2,21 +2,5 @@ if(WITH_PYTHON)
 cc_library(paddle_pybind SHARED
     SRCS pybind.cc
     DEPS pybind python backward
-    sgd_op
-    gather_op
-    scatter_op
-    add_op
-    mul_op
-    rowwise_add_op
-    sigmoid_op
-    softmax_op
-    mean_op
-    cross_entropy_op
-    recurrent_op
-    uniform_random_op
-    gaussian_random_op
-    fill_zeros_like_op
-    lookup_table_op
-    scale_op
-    minus_op)
+    ${GLOB_OP_LIB})
 endif(WITH_PYTHON)

proto/ModelConfig.proto

@@ -85,6 +85,12 @@ message ConvConfig {
 
   optional uint32 dilation = 15 [ default = 1 ];
   optional uint32 dilation_y = 16 [ default = 1 ];
+
+  optional uint32 filter_size_z = 17 [ default = 1 ];
+  optional uint32 padding_z = 18 [ default = 1 ];
+  optional uint32 stride_z = 19 [ default = 1 ];
+  optional uint32 output_z = 20 [ default = 1 ];
+  optional uint32 img_size_z = 21 [ default = 1 ];
 }
 
 message PoolConfig {
@@ -127,6 +133,12 @@ message PoolConfig {
 
   // if not set, use padding
   optional uint32 padding_y = 13;
+
+  optional uint32 size_z = 14 [ default = 1 ];
+  optional uint32 stride_z = 15 [ default = 1 ];
+  optional uint32 output_z = 16 [ default = 1 ];
+  optional uint32 img_size_z = 17 [ default = 1 ];
+  optional uint32 padding_z = 18 [ default = 1 ];
 }
 
 message SppConfig {
@@ -502,6 +514,8 @@ message LayerConfig {
 
   // for HuberRegressionLoss
   optional double delta = 57 [ default = 1.0 ];
+
+  optional uint64 depth = 58 [ default = 1 ];
 }
 
 message EvaluatorConfig {

python/paddle/trainer/config_parser.py

@@ -886,6 +886,36 @@ class Conv(Cfg):
             config_assert(output_x <= 0)
 
 
+# please refer to the comments in proto/ModelConfig.proto
+@config_class
+class Conv3D(Cfg):
+    def __init__(self,
+                 filter_size,
+                 channels,
+                 padding=None,
+                 stride=None,
+                 groups=None,
+                 filter_channels=None,
+                 output_x=None,
+                 img_size=None,
+                 caffe_mode=True,
+                 filter_size_y=None,
+                 padding_y=None,
+                 stride_y=None,
+                 filter_size_z=None,
+                 padding_z=None,
+                 stride_z=None):
+        self.add_keys(locals())
+        self.filter_size_y = filter_size_y if filter_size_y else filter_size
+        self.filter_size_z = filter_size_z if filter_size_z else filter_size
+        self.padding_y = padding_y if padding_y else padding
+        self.padding_z = padding_z if padding_z else padding
+        self.stride_y = stride_y if stride_y else stride
+        self.stride_z = stride_z if stride_z else stride
+        if output_x is not None:
+            config_assert(output_x <= 0)
+
+
 @config_class
 class BilinearInterp(Cfg):
     def __init__(self, out_size_x=None, out_size_y=None, channels=None):
@@ -908,6 +938,31 @@ class Pool(Cfg):
         self.add_keys(locals())
 
 
+@config_class
+class Pool3d(Cfg):
+    def __init__(
+            self,
+            pool_type,
+            channels,
+            size_x,
+            size_y=None,
+            size_z=None,
+            start=None,
+            stride=None,  # 1 by defalut in protobuf
+            stride_y=None,
+            stride_z=None,
+            padding=None,  # 0 by defalut in protobuf
+            padding_y=None,
+            padding_z=None):
+        self.add_keys(locals())
+        self.filter_size_y = size_y if size_y else size_x
+        self.filter_size_z = size_z if size_z else size_x
+        self.padding_y = padding_y if padding_y else padding
+        self.padding_z = padding_z if padding_z else padding
+        self.stride_y = stride_y if stride_y else stride
+        self.stride_z = stride_z if stride_z else stride
+
+
 @config_class
 class SpatialPyramidPool(Cfg):
     def __init__(self, pool_type, pyramid_height, channels):
@@ -1172,6 +1227,20 @@ def get_img_size(input_layer_name, channels):
     return img_size, img_size_y
 
 
+def get_img3d_size(input_layer_name, channels):
+    input = g_layer_map[input_layer_name]
+    img_pixels = input.size / channels
+    img_size = input.width
+    img_size_y = input.height
+    img_size_z = input.depth
+
+    config_assert(
+        img_size * img_size_y * img_size_z == img_pixels,
+        "Input layer %s: Incorrect input image size %d * %d * %d for input image pixels %d"
+        % (input_layer_name, img_size, img_size_y, img_size_z, img_pixels))
+    return img_size, img_size_y, img_size_z
+
+
 def parse_bilinear(bilinear, input_layer_name, bilinear_conf):
     parse_image(bilinear, input_layer_name, bilinear_conf.image_conf)
     bilinear_conf.out_size_x = bilinear.out_size_x
@@ -1209,6 +1278,45 @@ def parse_pool(pool, input_layer_name, pool_conf, ceil_mode):
                                          pool_conf.stride_y, not ceil_mode)
 
 
+def parse_pool3d(pool, input_layer_name, pool_conf, ceil_mode):
+    pool_conf.pool_type = pool.pool_type
+    config_assert(pool.pool_type in ['max-projection', 'avg-projection'],
+                  "pool-type %s is not in "
+                  "['max-projection', 'avg-projection']" % pool.pool_type)
+
+    pool_conf.channels = pool.channels
+
+    pool_conf.size_x = pool.size_x
+    pool_conf.stride = pool.stride
+    pool_conf.padding = pool.padding
+
+    pool_conf.size_y = default(pool.size_y, pool_conf.size_x)
+    pool_conf.size_z = default(pool.size_z, pool_conf.size_x)
+    pool_conf.stride_y = default(pool.stride_y, pool_conf.stride)
+    pool_conf.stride_z = default(pool.stride_z, pool_conf.stride)
+    pool_conf.padding_y = default(pool.padding_y, pool_conf.padding)
+    pool_conf.padding_z = default(pool.padding_z, pool_conf.padding)
+
+    pool_conf.img_size, pool_conf.img_size_y, pool_conf.img_size_z = \
+        get_img3d_size(input_layer_name, pool.channels)
+
+    config_assert(not pool.start, "start is deprecated in pooling.")
+
+    if pool.padding is not None:
+        pool_conf.padding = pool.padding
+    pool_conf.padding_y = default(pool.padding_y, pool_conf.padding)
+    pool_conf.padding_z = default(pool.padding_z, pool_conf.padding)
+    pool_conf.output_x = cnn_output_size(pool_conf.img_size, pool_conf.size_x,
+                                         pool_conf.padding, pool_conf.stride,
+                                         not ceil_mode)
+    pool_conf.output_y = cnn_output_size(pool_conf.img_size_y, pool_conf.size_y,
+                                         pool_conf.padding_y,
+                                         pool_conf.stride_y, not ceil_mode)
+    pool_conf.output_z = cnn_output_size(pool_conf.img_size_z, pool_conf.size_z,
+                                         pool_conf.padding_z,
+                                         pool_conf.stride_z, not ceil_mode)
+
+
 def parse_spp(spp, input_layer_name, spp_conf):
     parse_image(spp, input_layer_name, spp_conf.image_conf)
     spp_conf.pool_type = spp.pool_type
@@ -1282,6 +1390,50 @@ def parse_conv(conv, input_layer_name, conv_conf, num_filters, trans=False):
             conv_conf.stride_y, conv_conf.caffe_mode)
 
 
+#caffe_mode: compute the output size using floor instead of ceil,
+#            which is consistent of caffe and CuDNN's convention.
+def parse_conv3d(conv, input_layer_name, conv_conf, num_filters, trans=False):
+    conv_conf.filter_size = conv.filter_size
+    conv_conf.filter_size_y = conv.filter_size_y
+    conv_conf.filter_size_z = conv.filter_size_z
+    conv_conf.channels = conv.channels
+    conv_conf.padding = conv.padding
+    conv_conf.padding_y = conv.padding_y
+    conv_conf.padding_z = conv.padding_z
+    conv_conf.stride = conv.stride
+    conv_conf.stride_y = conv.stride_y
+    conv_conf.stride_z = conv.stride_z
+    conv_conf.groups = conv.groups
+    conv_conf.caffe_mode = conv.caffe_mode
+
+    if not trans:
+        conv_conf.filter_channels = conv.channels / conv.groups
+        conv_conf.img_size, conv_conf.img_size_y, conv_conf.img_size_z = \
+            get_img3d_size(input_layer_name, conv.channels)
+        conv_conf.output_x = cnn_output_size(
+            conv_conf.img_size, conv_conf.filter_size, conv_conf.padding,
+            conv_conf.stride, conv_conf.caffe_mode)
+        conv_conf.output_y = cnn_output_size(
+            conv_conf.img_size_y, conv_conf.filter_size_y, conv_conf.padding_y,
+            conv_conf.stride_y, conv_conf.caffe_mode)
+        conv_conf.output_z = cnn_output_size(
+            conv_conf.img_size_z, conv_conf.filter_size_z, conv_conf.padding_z,
+            conv_conf.stride_z, conv_conf.caffe_mode)
+    else:
+        conv_conf.filter_channels = num_filters / conv.groups
+        conv_conf.output_x, conv_conf.output_y, conv_conf.output_z = \
+            get_img3d_size(input_layer_name, conv.channels)
+        conv_conf.img_size = cnn_image_size(
+            conv_conf.output_x, conv_conf.filter_size, conv_conf.padding,
+            conv_conf.stride, conv_conf.caffe_mode)
+        conv_conf.img_size_y = cnn_image_size(
+            conv_conf.output_y, conv_conf.filter_size_y, conv_conf.padding_y,
+            conv_conf.stride_y, conv_conf.caffe_mode)
+        conv_conf.img_size_z = cnn_image_size(
+            conv_conf.output_z, conv_conf.filter_size_z, conv_conf.padding_z,
+            conv_conf.stride_z, conv_conf.caffe_mode)
+
+
 def parse_block_expand(block_expand, input_layer_name, block_expand_conf):
     block_expand_conf.channels = block_expand.channels
     block_expand_conf.stride_x = block_expand.stride_x
@@ -1585,6 +1737,9 @@ class LayerBase(object):
         self.config.height = height
         self.config.width = width
 
+    def set_layer_depth(self, depth):
+        self.config.depth = depth
+
     def set_cnn_layer(self,
                       input_layer_name,
                       height,
@@ -1607,6 +1762,21 @@ class MultiClassCrossEntropySelfNormCostLayer(LayerBase):
         self.config.softmax_selfnorm_alpha = softmax_selfnorm_alpha
 
 
+@config_layer('cross_entropy_over_beam')
+class CrossEntropyOverBeamLayer(LayerBase):
+    def __init__(self, name, inputs, **xargs):
+        config_assert(len(inputs) % 3 == 0, "Error input number.")
+        super(CrossEntropyOverBeamLayer, self).__init__(
+            name, 'cross_entropy_over_beam', 0, inputs, **xargs)
+        input_num = len(inputs) / 3
+        for i in range(input_num):
+            input_layer = self.get_input_layer(i * 3)
+            config_assert(input_layer.size == 1, (
+                "Inputs for this layer are made up of "
+                "several triples, in which the first one is scores over "
+                "all candidate paths, whose size should be equal to 1."))
+
+
 @config_layer('fc')
 class FCLayer(LayerBase):
     layer_type = 'fc'
@@ -1788,11 +1958,19 @@ class DetectionOutputLayer(LayerBase):
 
 @config_layer('data')
 class DataLayer(LayerBase):
-    def __init__(self, name, size, height=None, width=None, device=None):
+    def __init__(self,
+                 name,
+                 size,
+                 depth=None,
+                 height=None,
+                 width=None,
+                 device=None):
         super(DataLayer, self).__init__(
             name, 'data', size, inputs=[], device=device)
         if height and width:
             self.set_layer_height_width(height, width)
+        if depth:
+            self.set_layer_depth(depth)
 
 
 '''
@@ -1907,7 +2085,7 @@ class ConvLayerBase(LayerBase):
 
     def calc_parameter_size(self, conv_conf):
         return self.config.num_filters * conv_conf.filter_channels \
-                    * (conv_conf.filter_size * conv_conf.filter_size_y)
+               * (conv_conf.filter_size * conv_conf.filter_size_y)
 
 
 @config_layer('exconv')
@@ -1991,6 +2169,87 @@ class ConvTransLayer(ConvTransLayerBase):
     layer_type = 'cudnn_convt'
 
 
+@config_layer('conv_3d')
+class Conv3DLayerBase(LayerBase):
+    def __init__(self,
+                 name,
+                 inputs=[],
+                 bias=True,
+                 num_filters=None,
+                 shared_biases=True,
+                 **xargs):
+        super(Conv3DLayerBase, self).__init__(
+            name, self.layer_type, 0, inputs=inputs, **xargs)
+
+        if num_filters is not None:
+            self.config.num_filters = num_filters
+
+        # need to specify layer in config
+        self.config.type = self.layer_type
+
+        trans = False
+        if self.config.type == "deconv3d":
+            trans = True
+
+        if shared_biases is not None:
+            self.config.shared_biases = shared_biases
+
+        for input_index in xrange(len(self.inputs)):
+            input_layer = self.get_input_layer(input_index)
+            conv_conf = self.config.inputs[input_index].conv_conf
+            parse_conv3d(
+                self.inputs[input_index].conv,
+                input_layer.name,
+                conv_conf,
+                num_filters,
+                trans=trans
+            )  # for z-axis pad:0, strid:1, filter_size:1, img_size:1
+            psize = self.calc_parameter_size(conv_conf)
+            self.create_input_parameter(input_index, psize)
+            if trans:
+                self.set_cnn_layer(name, conv_conf.img_size_z,
+                                   conv_conf.img_size_y, conv_conf.img_size,
+                                   self.config.num_filters)
+            else:
+                self.set_cnn_layer(name, conv_conf.output_z, conv_conf.output_y,
+                                   conv_conf.output_x, self.config.num_filters)
+
+        psize = self.config.size
+        if shared_biases:
+            psize = self.config.num_filters
+        self.create_bias_parameter(bias, psize, [psize, 1])
+
+    def calc_parameter_size(self, conv_conf):
+        return self.config.num_filters * conv_conf.filter_channels \
+               * (conv_conf.filter_size * conv_conf.filter_size_y \
+                  * conv_conf.filter_size_z)
+
+    def set_cnn_layer(self,
+                      input_layer_name,
+                      depth,
+                      height,
+                      width,
+                      channels,
+                      is_print=True):
+        size = depth * height * width * channels
+        self.set_layer_size(size)
+        self.set_layer_height_width(height, width)
+        self.set_layer_depth(depth)
+        if is_print:
+            print("output for %s: c = %d, d = %d, h = %d, w = %d, size = %d" %
+                  (input_layer_name, channels, depth, height, width, size))
+
+
+@config_layer('conv3d')
+class Conv3DLayer(Conv3DLayerBase):
+    layer_type = 'conv3d'
+
+
+@config_layer('deconv3d')
+class Conv3DLayer(Conv3DLayerBase):
+    layer_type = 'deconv3d'
+
+
 @config_layer('norm')
 class NormLayer(LayerBase):
     def __init__(self, name, inputs, **xargs):
@@ -2020,6 +2279,35 @@ class PoolLayer(LayerBase):
                                pool_conf.channels)
 
 
+@config_layer('pool3d')
+class Pool3DLayer(LayerBase):
+    def __init__(self, name, inputs, ceil_mode=True, **xargs):
+        super(Pool3DLayer, self).__init__(
+            name, 'pool3d', 0, inputs=inputs, **xargs)
+        for input_index in xrange(len(self.inputs)):
+            input_layer = self.get_input_layer(input_index)
+            pool_conf = self.config.inputs[input_index].pool_conf
+            parse_pool3d(self.inputs[input_index].pool, input_layer.name,
+                         pool_conf, ceil_mode)
+            self.set_cnn_layer(name, pool_conf.output_z, pool_conf.output_y,
+                               pool_conf.output_x, pool_conf.channels)
+
+    def set_cnn_layer(self,
+                      input_layer_name,
+                      depth,
+                      height,
+                      width,
+                      channels,
+                      is_print=True):
+        size = depth * height * width * channels
+        self.set_layer_size(size)
+        self.set_layer_height_width(height, width)
+        self.set_layer_depth(depth)
+        if is_print:
+            print("output for %s: c = %d, d = %d, h = %d, w = %d, size = %d" %
+                  (input_layer_name, channels, depth, height, width, size))
+
+
 @config_layer('spp')
 class SpatialPyramidPoolLayer(LayerBase):
     def __init__(self, name, inputs, **xargs):
@@ -2268,6 +2556,7 @@ def define_cost(class_name, cost_type):
 
 
 define_cost('MultiClassCrossEntropy', 'multi-class-cross-entropy')
+define_cost('CrossEntropyOverBeamCostLayer', 'cross_entropy_over_beam')
 define_cost('RankingCost', 'rank-cost')
 define_cost('AucValidation', 'auc-validation')
 define_cost('PnpairValidation', 'pnpair-validation')

python/paddle/trainer_config_helpers/layers.py

@@ -11,7 +11,6 @@
 # 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.
-
 import functools
 import collections
 import inspect
@@ -106,6 +105,8 @@ __all__ = [
     'nce_layer',
     'cross_entropy_with_selfnorm',
     'cross_entropy',
+    'BeamInput',
+    'cross_entropy_over_beam',
     'multi_binary_label_cross_entropy',
     'sum_cost',
     'rank_cost',
@@ -136,8 +137,10 @@ __all__ = [
     'clip_layer',
     'slice_projection',
     'seq_slice_layer',
-    'kmax_sequence_score_layer',
+    'kmax_seq_score_layer',
+    'img_pool3d_layer',
     'scale_shift_layer',
+    'img_conv3d_layer',
 ]
 
 
@@ -166,6 +169,7 @@ class LayerType(object):
     EXCONVTRANS_LAYER = 'exconvt'
     CUDNNCONV_LAYER = 'cudnn_conv'
     POOL_LAYER = 'pool'
+    POOL3D_LAYER = 'pool3d'
     BATCH_NORM_LAYER = 'batch_norm'
     NORM_LAYER = 'norm'
     SUM_TO_ONE_NORM_LAYER = 'sum_to_one_norm'
@@ -219,12 +223,16 @@ class LayerType(object):
     CRF_DECODING_LAYER = 'crf_decoding'
     NCE_LAYER = 'nce'
 
+    CONV3D_LAYER = 'conv3d'
+    DECONV3D_LAYER = 'deconv3d'
+
     RANK_COST = 'rank-cost'
     LAMBDA_COST = 'lambda_cost'
     HUBER_REGRESSION = 'huber_regression'
     HUBER_CLASSIFICATION = 'huber_classification'
     CROSS_ENTROPY = 'multi-class-cross-entropy'
     CROSS_ENTROPY_WITH_SELFNORM = 'multi_class_cross_entropy_with_selfnorm'
+    CROSS_ENTROPY_OVER_BEAM = 'cross_entropy_over_beam'
     SOFT_BIN_CLASS_CROSS_ENTROPY = 'soft_binary_class_cross_entropy'
     MULTI_BIN_LABEL_CROSS_ENTROPY = 'multi_binary_label_cross_entropy'
     SUM_COST = 'sum_cost'
@@ -894,7 +902,8 @@ def mixed_layer(size=0,
 
 
 @layer_support()
-def data_layer(name, size, height=None, width=None, layer_attr=None):
+def data_layer(name, size, depth=None, height=None, width=None,
+               layer_attr=None):
     """
     Define DataLayer For NeuralNetwork.
 
@@ -921,15 +930,18 @@ def data_layer(name, size, height=None, width=None, layer_attr=None):
         type=LayerType.DATA,
         name=name,
         size=size,
+        depth=depth,
         height=height,
         width=width,
         **ExtraLayerAttribute.to_kwargs(layer_attr))
 
+    if depth is None:
+        depth = 1
     num_filters = None
     if height is not None and width is not None:
-        num_filters = size / (width * height)
-        assert num_filters * width * height == size, \
-            "size=%s width=%s height=%s" % (size, width, height)
+        num_filters = size / (width * height * depth)
+        assert num_filters * width * height * depth == size, \
+                "size=%s width=%s height=%s depth=%s"  % (size, width, height, depth)
 
     return LayerOutput(name, LayerType.DATA, size=size, num_filters=num_filters)
 
@@ -2653,6 +2665,146 @@ def img_pool_layer(input,
         size=l.config.size)
 
 
+@wrap_name_default("pool3d")
+@layer_support()
+def img_pool3d_layer(input,
+                     pool_size,
+                     name=None,
+                     num_channels=None,
+                     pool_type=None,
+                     stride=1,
+                     padding=0,
+                     layer_attr=None,
+                     pool_size_y=None,
+                     stride_y=None,
+                     padding_y=None,
+                     pool_size_z=None,
+                     stride_z=None,
+                     padding_z=None,
+                     ceil_mode=True):
+    """
+    Image pooling Layer.
+
+    The details of pooling layer, please refer ufldl's pooling_ .
+
+    .. _pooling: http://ufldl.stanford.edu/tutorial/supervised/Pooling/
+
+    - ceil_mode=True:
+
+    ..  math::
+
+        w = 1 + int(ceil(input\_width + 2 * padding - pool\_size) / float(stride))
+        h = 1 + int(ceil(input\_height + 2 * padding\_y - pool\_size\_y) / float(stride\_y))
+        d = 1 + int(ceil(input\_depth + 2 * padding\_z - pool\_size\_z) / float(stride\_z))
+
+    - ceil_mode=False:
+
+    ..  math::
+
+        w = 1 + int(floor(input\_width + 2 * padding - pool\_size) / float(stride))
+        h = 1 + int(floor(input\_height + 2 * padding\_y - pool\_size\_y) / float(stride\_y))
+        d = 1 + int(floor(input\_depth + 2 * padding\_z - pool\_size\_z) / float(stride\_z))
+
+    The example usage is:
+
+    ..  code-block:: python
+
+        maxpool = img_pool3d_layer(input=conv,
+                                 pool_size=3,
+                                 num_channels=8,
+                                 stride=1,
+                                 padding=1,
+                                 pool_type=MaxPooling())
+
+    :param padding: pooling padding width.
+    :type padding: int|tuple|list
+    :param name: name of pooling layer
+    :type name: basestring.
+    :param input: layer's input
+    :type input: LayerOutput
+    :param pool_size: pooling window width
+    :type pool_size: int|tuple|list
+    :param num_channels: number of input channel.
+    :type num_channels: int
+    :param pool_type: pooling type. MaxPooling or AvgPooling. Default is
+                      MaxPooling.
+    :type pool_type: BasePoolingType
+    :param stride: stride width of pooling.
+    :type stride: int|tuple|list
+    :param layer_attr: Extra Layer attribute.
+    :type layer_attr: ExtraLayerAttribute
+    :param ceil_mode: Wether to use ceil mode to calculate output height and with.
+                      Defalut is True. If set false, Otherwise use floor.
+
+    :type ceil_mode: bool
+    :return: LayerOutput object.
+    :rtype: LayerOutput
+    """
+    if num_channels is None:
+        assert input.num_filters is not None
+        num_channels = input.num_filters
+
+    if pool_type is None:
+        pool_type = MaxPooling()
+    elif isinstance(pool_type, AvgPooling):
+        pool_type.name = 'avg'
+
+    type_name = pool_type.name + '-projection' \
+        if (
+        isinstance(pool_type, AvgPooling) or isinstance(pool_type, MaxPooling)) \
+        else pool_type.name
+
+    if isinstance(pool_size, collections.Sequence):
+        assert len(pool_size) == 3
+        pool_size, pool_size_y, pool_size_z = pool_size
+    else:
+        pool_size_y = pool_size
+        pool_size_z = pool_size
+
+    if isinstance(stride, collections.Sequence):
+        assert len(stride) == 3
+        stride, stride_y, stride_z = stride
+    else:
+        stride_y = stride
+        stride_z = stride
+
+    if isinstance(padding, collections.Sequence):
+        assert len(padding) == 3
+        padding, padding_y, padding_y = padding
+    else:
+        padding_y = padding
+        padding_z = padding
+
+    l = Layer(
+        name=name,
+        type=LayerType.POOL3D_LAYER,
+        inputs=[
+            Input(
+                input.name,
+                pool=Pool3d(
+                    pool_type=type_name,
+                    channels=num_channels,
+                    size_x=pool_size,
+                    start=None,
+                    stride=stride,
+                    padding=padding,
+                    size_y=pool_size_y,
+                    stride_y=stride_y,
+                    padding_y=padding_y,
+                    size_z=pool_size_z,
+                    stride_z=stride_z,
+                    padding_z=padding_z))
+        ],
+        ceil_mode=ceil_mode,
+        **ExtraLayerAttribute.to_kwargs(layer_attr))
+    return LayerOutput(
+        name,
+        LayerType.POOL_LAYER,
+        parents=[input],
+        num_filters=num_channels,
+        size=l.config.size)
+
+
 @wrap_name_default("spp")
 @layer_support()
 def spp_layer(input,
@@ -4071,8 +4223,12 @@ def __cost_input__(input, label, weight=None):
     """
     inputs and parents for cost layers.
     """
-    ipts = [Input(input.name), Input(label.name)]
-    parents = [input, label]
+    if isinstance(input, LayerOutput):
+        input = [input]
+    if isinstance(label, LayerOutput):
+        label = [label]
+    ipts = [Input(ipt.name) for ipt in (input + label)]
+    parents = [ipt for ipt in (input + label)]
     if weight is not None:
         assert weight.size == 1
         ipts.append(Input(weight.name))
@@ -5059,17 +5215,6 @@ def warp_ctc_layer(input,
     building process, PaddlePaddle will clone the source codes, build and
     install it to :code:`third_party/install/warpctc` directory.
 
-    To use warp_ctc layer, you need to specify the path of :code:`libwarpctc.so`,
-    using following methods:
-
-    1. Set it in :code:`paddle.init` (python api) or :code:`paddle_init` (c api),
-    such as :code:`paddle.init(use_gpu=True,
-    warpctc_dir=your_paddle_source_dir/third_party/install/warpctc/lib)`.
-
-    2. Set environment variable LD_LIBRARY_PATH on Linux or DYLD_LIBRARY_PATH
-    on Mac OS. For instance, :code:`export
-    LD_LIBRARY_PATH=your_paddle_source_dir/third_party/install/warpctc/lib:$LD_LIBRARY_PATH`.
-
     More details of CTC can be found by referring to `Connectionist Temporal
     Classification: Labelling Unsegmented Sequence Data with Recurrent
     Neural Networks <http://machinelearning.wustl.edu/mlpapers/paper_files/
@@ -5777,10 +5922,10 @@ def multi_binary_label_cross_entropy(input,
 
     if input.activation is None or \
             not isinstance(input.activation, SigmoidActivation):
-        logger.log(
-            logging.WARN,
-            "%s is not recommend for multi_binary_label_cross_entropy's activation, "
-            "maybe the sigmoid is better" % repr(input.activation))
+        logger.log(logging.WARN,
+                   ("%s is not a recommended activation for "
+                    "multi_binary_label_cross_entropy, sigmoid is better") %
+                   repr(input.activation))
 
     Layer(
         name=name,
@@ -5795,6 +5940,113 @@ def multi_binary_label_cross_entropy(input,
         size=1)
 
 
+class BeamInput(object):
+    """
+    Define the input for cross_entropy_over_beam layer.
+
+    A beam is made up of a triple: the first one is scores over all
+    candidates; the second one is indices of top k selected candidates; the
+    third one is the index of ground truth, which is also always called
+    gold.
+    """
+
+    def __init__(self, candidate_scores, selected_candidates, gold):
+        assert isinstance(candidate_scores, LayerOutput)
+        self.candidate_scores = candidate_scores
+        assert candidate_scores.size == 1
+
+        assert isinstance(selected_candidates, LayerOutput)
+        self.selected_candidates = selected_candidates
+
+        assert isinstance(gold, LayerOutput)
+        self.gold = gold
+
+
+@wrap_name_default()
+@layer_support()
+def cross_entropy_over_beam(input, name=None):
+    """
+    This layer is used in learning to search models, which is to solve complex
+    joint prediction problems based on learning to search through a
+    problem-defined search space.
+
+    Specifically, the learning to search process for this layer begins with
+    searching a target sequence from a nested sequence. In the first search
+    step, top beam size sequences with highest scores, indices of these top k
+    sequences in the original nested sequence, and the ground truth (also
+    called gold) altogether (a triple) make up of the first beam.
+
+    Then, several special positions, for example, start and end positions
+    that define meaningful segments are searched. In these searches, top k
+    positions with highest scores are selected, and then sequence, starting
+    from the selected starts till ends of the sequences (or a fixed position)
+    are taken to search next.
+
+    We call the possible top k results returned in one search the beam. This
+    search process can be repeated for pre-defined turns and leads to several
+    beam expansions.
+
+    Finally, the layer cross_entropy_over_beam takes all the beam expansions
+    which contain several candidate targets found along the multi-step search.
+    cross_entropy_over_beam calculates cross entropy over the expanded beams
+    which all the candidates in the beam as the normalized factor.
+
+    Note that, if gold falls off the beam at search step t, then the cost is
+    calculated over the beam at step t.
+
+    This cost layer always works together with kmax_seq_score_layer,
+    sub_nested_seq_layer, and sequence_slice_layer to trim the input to form a
+    sub-search space.
+
+
+    The example usage is:
+
+    .. code-block:: python
+
+       cost = cross_entropy_over_beam(input=[
+           BeamInput(
+               candidate_scores=beam1_candidates,
+               selected_candidates=beam1_topk,
+               gold=gold1),
+           BeamInput(
+               candidate_scores=beam2_candidates,
+               selected_candidates=beam2_topk,
+               gold=gold2),
+       ])
+
+
+    :param input: input beams for this layer.
+    :type input: BeamInput
+    :param name: input beams for this layer.
+    :type name: basestring
+    :return: LayerOutput object.
+    :rtype: LayerOutput
+    """
+
+    if isinstance(input, BeamInput):
+        input = [input]
+    else:
+        assert isinstance(input, list), (
+            'input for cross_entropy_over_beam shold be a python list '
+            'of BeamInput object.')
+        for ipt in input:
+            assert isinstance(ipt, BeamInput), (
+                'input for cross_entropy_over_beam '
+                'should be a BeamInput object.')
+
+    ipts = []
+    parents = []
+    for beam in input:
+        parents += [beam.candidate_scores, beam.selected_candidates, beam.gold]
+        ipts += [
+            beam.candidate_scores.name, beam.selected_candidates.name,
+            beam.gold.name
+        ]
+
+    Layer(name=name, type=LayerType.CROSS_ENTROPY_OVER_BEAM, inputs=ipts)
+    return LayerOutput(name, LayerType.CROSS_ENTROPY, parents=parents, size=1)
+
+
 @wrap_name_default()
 @layer_support()
 def smooth_l1_cost(input, label, name=None, coeff=1.0, layer_attr=None):
@@ -6345,14 +6597,14 @@ def seq_slice_layer(input, starts, ends, name=None):
 
 @wrap_name_default()
 @layer_support()
-def kmax_sequence_score_layer(input, name=None, beam_size=1):
+def kmax_seq_score_layer(input, name=None, beam_size=1):
     """
     This layer accepts one input which are scores over a sequence or a nested
     sequence, and returns indices of beam_size sequences with highest scores.
 
     .. code-block:: python
 
-        kmax_indices = kmax_sequence_score_layer(input=input_layer, beam_size)
+        kmax_indices = kmax_seq_score_layer(input=input_layer, beam_size)
 
 
     :param name: The Layer Name.
@@ -6365,10 +6617,10 @@ def kmax_sequence_score_layer(input, name=None, beam_size=1):
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
-    assert isinstance(input, LayerOutput), ("kmax_sequence_score_layer "
+    assert isinstance(input, LayerOutput), ("kmax_seq_score_layer "
                                             "accepts only one input.")
     assert input.size == 1, (
-        "input of kmax_sequence_score_layer is a score"
+        "input of kmax_seq_score_layer is a score "
         "over a sequence or a nested sequence, so its width must be 1.")
 
     Layer(
@@ -6381,6 +6633,149 @@ def kmax_sequence_score_layer(input, name=None, beam_size=1):
         name, LayerType.KMAX_SEQ_SCORE, parents=[input], size=input.size)
 
 
+@wrap_name_default("conv3d")
+@wrap_param_attr_default()
+@wrap_bias_attr_default()
+@wrap_act_default(act=ReluActivation())
+@layer_support(DROPOUT)
+def img_conv3d_layer(input,
+                     filter_size,
+                     num_filters,
+                     name=None,
+                     num_channels=None,
+                     act=None,
+                     groups=1,
+                     stride=1,
+                     padding=0,
+                     bias_attr=None,
+                     param_attr=None,
+                     shared_biases=True,
+                     layer_attr=None,
+                     trans=False,
+                     layer_type=None):
+    """
+
+    The example usage is:
+
+    ..  code-block:: python
+
+        conv = img_conv3d_layer(input=data, filter_size=1,
+                              num_channels=8,
+                              num_filters=16, stride=1,
+                              bias_attr=False,
+                              act=ReluActivation())
+
+    :param name: Layer name.
+    :type name: basestring
+    :param input: Layer Input.
+    :type input: LayerOutput
+    :param filter_size: The x dimension of a filter kernel. Or input a list.
+    :type filter_size: int|tuple|list
+    :param num_filters: Each filter group's number of filter
+    :param act: Activation type. Default is tanh
+    :type act: BaseActivation
+    :param groups: Group size of filters.
+    :type groups: int
+    :param stride: The x dimension of the stride. Or input a tuple for two image
+                   dimension.
+    :type stride: int|tuple|list
+    :param padding: The x dimension of the padding. Or input a tuple for two
+                    image dimension
+    :type padding: int|tuple|list
+    :param bias_attr: Convolution bias attribute. None means default bias.
+                      False means no bias.
+    :type bias_attr: ParameterAttribute|False
+    :param num_channels: number of input channels. If None will be set
+                        automatically from previous output.
+    :type num_channels: int
+    :param param_attr: Convolution param attribute. None means default attribute
+    :type param_attr: ParameterAttribute
+    :param shared_biases: Is biases will be shared between filters or not.
+    :type shared_biases: bool
+    :param layer_attr: Layer Extra Attribute.
+    :type layer_attr: ExtraLayerAttribute
+    :param trans: true if it is a convTransLayer, false if it is a convLayer
+    :type trans: bool
+    :param layer_type: specify the layer_type, default is None. If trans=True,
+                       layer_type has to be "exconvt" or "cudnn_convt",
+                       otherwise layer_type has to be either "exconv" or
+                       "cudnn_conv"
+    :type layer_type: String
+    :return: LayerOutput object.
+    :rtype: LayerOutput
+    """
+    if num_channels is None:
+        assert input.num_filters is not None
+        num_channels = input.num_filters
+
+    if isinstance(filter_size, collections.Sequence):
+        assert len(filter_size) == 3
+        filter_size, filter_size_y, filter_size_z = filter_size
+    else:
+        filter_size_y = filter_size
+        filter_size_z = filter_size
+
+    if isinstance(stride, collections.Sequence):
+        assert len(stride) == 3
+        stride, stride_y, stride_z = stride
+    else:
+        stride_y = stride
+        stride_z = stride
+
+    if isinstance(padding, collections.Sequence):
+        assert len(padding) == 3
+        padding, padding_y, padding_z = padding
+    else:
+        padding_y = padding
+        padding_z = padding
+
+    if param_attr.attr.get('initial_smart'):
+        # special initial for conv layers.
+        init_w = (2.0 / (filter_size**2 * num_channels))**0.5
+        param_attr.attr["initial_mean"] = 0.0
+        param_attr.attr["initial_std"] = init_w
+        param_attr.attr["initial_strategy"] = 0
+        param_attr.attr["initial_smart"] = False
+
+    if layer_type:
+        if trans:
+            assert layer_type in ["deconv3d"]
+        lt = layer_type
+    else:
+        lt = LayerType.DECONV3D_LAYER if trans else LayerType.CONV3D_LAYER
+
+    l = Layer(
+        name=name,
+        inputs=Input(
+            input.name,
+            conv=Conv3D(
+                filter_size=filter_size,
+                padding=padding,
+                stride=stride,
+                channels=num_channels,
+                groups=groups,
+                filter_size_y=filter_size_y,
+                padding_y=padding_y,
+                stride_y=stride_y,
+                filter_size_z=filter_size_z,
+                padding_z=padding_z,
+                stride_z=stride_z),
+            **param_attr.attr),
+        active_type=act.name,
+        num_filters=num_filters,
+        bias=ParamAttr.to_bias(bias_attr),
+        shared_biases=shared_biases,
+        type=lt,
+        **ExtraLayerAttribute.to_kwargs(layer_attr))
+    return LayerOutput(
+        name,
+        lt,
+        parents=[input],
+        activation=act,
+        num_filters=num_filters,
+        size=l.config.size)
+
+
 @wrap_name_default("scale_shift")
 @wrap_param_attr_default()
 @wrap_bias_attr_default()

python/paddle/trainer_config_helpers/tests/configs/file_list.sh

@@ -8,7 +8,8 @@ test_spp_layer test_bilinear_interp test_maxout test_bi_grumemory math_ops
 test_seq_concat_reshape test_pad test_smooth_l1 test_multiplex_layer
 test_prelu_layer test_row_conv test_detection_output_layer test_multibox_loss_layer
 test_recursive_topology test_gated_unit_layer test_clip_layer test_row_l2_norm_layer
-test_kmax_seq_socre_layer test_seq_select_layers test_scale_shift_layer
-test_seq_slice_layer)
+test_kmax_seq_socre_layer test_sub_nested_seq_select_layer test_scale_shift_layer
+test_seq_slice_layer test_cross_entropy_over_beam test_pooling3D_layer
+test_conv3d_layer test_deconv3d_layer)
 
 export whole_configs=(test_split_datasource)

python/paddle/trainer_config_helpers/tests/configs/protostr/test_conv3d_layer.protostr

+type: "nn"
+layers {
+  name: "data"
+  type: "data"
+  size: 36288
+  active_type: ""
+  height: 48
+  width: 42
+  depth: 6
+}
+layers {
+  name: "conv3d_1"
+  type: "conv3d"
+  size: 24192
+  active_type: ""
+  inputs {
+    input_layer_name: "data"
+    input_parameter_name: "_conv3d_1.w0"
+    conv_conf {
+      filter_size: 3
+      channels: 3
+      stride: 2
+      padding: 1
+      groups: 1
+      filter_channels: 3
+      output_x: 21
+      img_size: 42
+      caffe_mode: true
+      filter_size_y: 3
+      padding_y: 1
+      stride_y: 2
+      output_y: 24
+      img_size_y: 48
+      filter_size_z: 3
+      padding_z: 1
+      stride_z: 2
+      output_z: 3
+      img_size_z: 6
+    }
+  }
+  bias_parameter_name: "_conv3d_1.wbias"
+  num_filters: 16
+  shared_biases: true
+  height: 24
+  width: 21
+  depth: 3
+}
+layers {
+  name: "conv3d_2"
+  type: "conv3d"
+  size: 24192
+  active_type: ""
+  inputs {
+    input_layer_name: "data"
+    input_parameter_name: "_conv3d_2.w0"
+    conv_conf {
+      filter_size: 3
+      channels: 3
+      stride: 2
+      padding: 1
+      groups: 1
+      filter_channels: 3
+      output_x: 21
+      img_size: 42
+      caffe_mode: true
+      filter_size_y: 3
+      padding_y: 1
+      stride_y: 2
+      output_y: 24
+      img_size_y: 48
+      filter_size_z: 3
+      padding_z: 1
+      stride_z: 2
+      output_z: 3
+      img_size_z: 6
+    }
+  }
+  bias_parameter_name: "_conv3d_2.wbias"
+  num_filters: 16
+  shared_biases: true
+  height: 24
+  width: 21
+  depth: 3
+}
+parameters {
+  name: "_conv3d_1.w0"
+  size: 1296
+  initial_mean: 0.0
+  initial_std: 0.272165526976
+  initial_strategy: 0
+  initial_smart: false
+}
+parameters {
+  name: "_conv3d_1.wbias"
+  size: 16
+  initial_mean: 0.0
+  initial_std: 0.0
+  dims: 16
+  dims: 1
+  initial_strategy: 0
+  initial_smart: false
+}
+parameters {
+  name: "_conv3d_2.w0"
+  size: 1296
+  initial_mean: 0.0
+  initial_std: 0.272165526976
+  initial_strategy: 0
+  initial_smart: false
+}
+parameters {
+  name: "_conv3d_2.wbias"
+  size: 16
+  initial_mean: 0.0
+  initial_std: 0.0
+  dims: 16
+  dims: 1
+  initial_strategy: 0
+  initial_smart: false
+}
+input_layer_names: "data"
+output_layer_names: "conv3d_2"
+sub_models {
+  name: "root"
+  layer_names: "data"
+  layer_names: "conv3d_1"
+  layer_names: "conv3d_2"
+  input_layer_names: "data"
+  output_layer_names: "conv3d_2"
+  is_recurrent_layer_group: false
+}
+

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

+type: "nn"
+layers {
+  name: "sentence_states"
+  type: "data"
+  size: 32
+  active_type: ""
+}
+layers {
+  name: "sentence_scores"
+  type: "data"
+  size: 1
+  active_type: ""
+}
+layers {
+  name: "__kmax_seq_score_layer_0__"
+  type: "kmax_seq_score"
+  active_type: ""
+  inputs {
+    input_layer_name: "sentence_scores"
+  }
+  beam_size: 5
+}
+layers {
+  name: "__sub_nested_seq_layer_0__"
+  type: "sub_nested_seq"
+  size: 32
+  active_type: ""
+  inputs {
+    input_layer_name: "sentence_states"
+  }
+  inputs {
+    input_layer_name: "__kmax_seq_score_layer_0__"
+  }
+}
+layers {
+  name: "__fc_layer_0__"
+  type: "fc"
+  size: 1
+  active_type: ""
+  inputs {
+    input_layer_name: "__sub_nested_seq_layer_0__"
+    input_parameter_name: "___fc_layer_0__.w0"
+  }
+  bias_parameter_name: "___fc_layer_0__.wbias"
+}
+layers {
+  name: "__kmax_seq_score_layer_1__"
+  type: "kmax_seq_score"
+  active_type: ""
+  inputs {
+    input_layer_name: "sentence_scores"
+  }
+  beam_size: 5
+}
+layers {
+  name: "__seq_slice_layer_0__"
+  type: "seq_slice"
+  size: 32
+  active_type: ""
+  inputs {
+    input_layer_name: "__sub_nested_seq_layer_0__"
+  }
+  inputs {
+    input_layer_name: "__kmax_seq_score_layer_1__"
+  }
+  select_first: true
+}
+layers {
+  name: "__fc_layer_1__"
+  type: "fc"
+  size: 1
+  active_type: ""
+  inputs {
+    input_layer_name: "__seq_slice_layer_0__"
+    input_parameter_name: "___fc_layer_1__.w0"
+  }
+  bias_parameter_name: "___fc_layer_1__.wbias"
+}
+layers {
+  name: "__kmax_seq_score_layer_2__"
+  type: "kmax_seq_score"
+  active_type: ""
+  inputs {
+    input_layer_name: "__fc_layer_1__"
+  }
+  beam_size: 5
+}
+layers {
+  name: "sentences_ids"
+  type: "data"
+  size: 1
+  active_type: ""
+}
+layers {
+  name: "start_ids"
+  type: "data"
+  size: 1
+  active_type: ""
+}
+layers {
+  name: "end_ids"
+  type: "data"
+  size: 1
+  active_type: ""
+}
+layers {
+  name: "__cross_entropy_over_beam_0__"
+  type: "cross_entropy_over_beam"
+  active_type: ""
+  inputs {
+    input_layer_name: "sentence_scores"
+  }
+  inputs {
+    input_layer_name: "__kmax_seq_score_layer_0__"
+  }
+  inputs {
+    input_layer_name: "sentences_ids"
+  }
+  inputs {
+    input_layer_name: "__fc_layer_0__"
+  }
+  inputs {
+    input_layer_name: "__kmax_seq_score_layer_1__"
+  }
+  inputs {
+    input_layer_name: "start_ids"
+  }
+  inputs {
+    input_layer_name: "__fc_layer_1__"
+  }
+  inputs {
+    input_layer_name: "__kmax_seq_score_layer_2__"
+  }
+  inputs {
+    input_layer_name: "end_ids"
+  }
+}
+parameters {
+  name: "___fc_layer_0__.w0"
+  size: 32
+  initial_mean: 0.0
+  initial_std: 0.176776695297
+  dims: 32
+  dims: 1
+  initial_strategy: 0
+  initial_smart: true
+}
+parameters {
+  name: "___fc_layer_0__.wbias"
+  size: 1
+  initial_mean: 0.0
+  initial_std: 0.0
+  dims: 1
+  dims: 1
+  initial_strategy: 0
+  initial_smart: false
+}
+parameters {
+  name: "___fc_layer_1__.w0"
+  size: 32
+  initial_mean: 0.0
+  initial_std: 0.176776695297
+  dims: 32
+  dims: 1
+  initial_strategy: 0
+  initial_smart: true
+}
+parameters {
+  name: "___fc_layer_1__.wbias"
+  size: 1
+  initial_mean: 0.0
+  initial_std: 0.0
+  dims: 1
+  dims: 1
+  initial_strategy: 0
+  initial_smart: false
+}
+input_layer_names: "sentence_scores"
+input_layer_names: "sentences_ids"
+input_layer_names: "sentence_states"
+input_layer_names: "start_ids"
+input_layer_names: "end_ids"
+output_layer_names: "__cross_entropy_over_beam_0__"
+sub_models {
+  name: "root"
+  layer_names: "sentence_states"
+  layer_names: "sentence_scores"
+  layer_names: "__kmax_seq_score_layer_0__"
+  layer_names: "__sub_nested_seq_layer_0__"
+  layer_names: "__fc_layer_0__"
+  layer_names: "__kmax_seq_score_layer_1__"
+  layer_names: "__seq_slice_layer_0__"
+  layer_names: "__fc_layer_1__"
+  layer_names: "__kmax_seq_score_layer_2__"
+  layer_names: "sentences_ids"
+  layer_names: "start_ids"
+  layer_names: "end_ids"
+  layer_names: "__cross_entropy_over_beam_0__"
+  input_layer_names: "sentence_scores"
+  input_layer_names: "sentences_ids"
+  input_layer_names: "sentence_states"
+  input_layer_names: "start_ids"
+  input_layer_names: "end_ids"
+  output_layer_names: "__cross_entropy_over_beam_0__"
+  is_recurrent_layer_group: false
+}
+

python/paddle/trainer_config_helpers/tests/configs/protostr/test_deconv3d_layer.protostr

+type: "nn"
+layers {
+  name: "data"
+  type: "data"
+  size: 36288
+  active_type: ""
+  height: 48
+  width: 42
+  depth: 6
+}
+layers {
+  name: "deconv3d_1"
+  type: "deconv3d"
+  size: 1387760
+  active_type: ""
+  inputs {
+    input_layer_name: "data"
+    input_parameter_name: "_deconv3d_1.w0"
+    conv_conf {
+      filter_size: 3
+      channels: 3
+      stride: 2
+      padding: 1
+      groups: 1
+      filter_channels: 16
+      output_x: 42
+      img_size: 83
+      caffe_mode: true
+      filter_size_y: 3
+      padding_y: 1
+      stride_y: 2
+      output_y: 48
+      img_size_y: 95
+      filter_size_z: 3
+      padding_z: 1
+      stride_z: 2
+      output_z: 6
+      img_size_z: 11
+    }
+  }
+  bias_parameter_name: "_deconv3d_1.wbias"
+  num_filters: 16
+  shared_biases: true
+  height: 95
+  width: 83
+  depth: 11
+}
+layers {
+  name: "deconv3d_2"
+  type: "deconv3d"
+  size: 1387760
+  active_type: ""
+  inputs {
+    input_layer_name: "data"
+    input_parameter_name: "_deconv3d_2.w0"
+    conv_conf {
+      filter_size: 3
+      channels: 3
+      stride: 2
+      padding: 1
+      groups: 1
+      filter_channels: 16
+      output_x: 42
+      img_size: 83
+      caffe_mode: true
+      filter_size_y: 3
+      padding_y: 1
+      stride_y: 2
+      output_y: 48
+      img_size_y: 95
+      filter_size_z: 3
+      padding_z: 1
+      stride_z: 2
+      output_z: 6
+      img_size_z: 11
+    }
+  }
+  bias_parameter_name: "_deconv3d_2.wbias"
+  num_filters: 16
+  shared_biases: true
+  height: 95
+  width: 83
+  depth: 11
+}
+parameters {
+  name: "_deconv3d_1.w0"
+  size: 6912
+  initial_mean: 0.0
+  initial_std: 0.272165526976
+  initial_strategy: 0
+  initial_smart: false
+}
+parameters {
+  name: "_deconv3d_1.wbias"
+  size: 16
+  initial_mean: 0.0
+  initial_std: 0.0
+  dims: 16
+  dims: 1
+  initial_strategy: 0
+  initial_smart: false
+}
+parameters {
+  name: "_deconv3d_2.w0"
+  size: 6912
+  initial_mean: 0.0
+  initial_std: 0.272165526976
+  initial_strategy: 0
+  initial_smart: false
+}
+parameters {
+  name: "_deconv3d_2.wbias"
+  size: 16
+  initial_mean: 0.0
+  initial_std: 0.0
+  dims: 16
+  dims: 1
+  initial_strategy: 0
+  initial_smart: false
+}
+input_layer_names: "data"
+output_layer_names: "deconv3d_2"
+sub_models {
+  name: "root"
+  layer_names: "data"
+  layer_names: "deconv3d_1"
+  layer_names: "deconv3d_2"
+  input_layer_names: "data"
+  output_layer_names: "deconv3d_2"
+  is_recurrent_layer_group: false
+}
+

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

@@ -17,7 +17,7 @@ layers {
   bias_parameter_name: "___fc_layer_0__.wbias"
 }
 layers {
-  name: "__kmax_sequence_score_layer_0__"
+  name: "__kmax_seq_score_layer_0__"
   type: "kmax_seq_score"
   active_type: ""
   inputs {
@@ -46,14 +46,14 @@ parameters {
   initial_smart: false
 }
 input_layer_names: "input_seq"
-output_layer_names: "__kmax_sequence_score_layer_0__"
+output_layer_names: "__kmax_seq_score_layer_0__"
 sub_models {
   name: "root"
   layer_names: "input_seq"
   layer_names: "__fc_layer_0__"
-  layer_names: "__kmax_sequence_score_layer_0__"
+  layer_names: "__kmax_seq_score_layer_0__"
   input_layer_names: "input_seq"
-  output_layer_names: "__kmax_sequence_score_layer_0__"
+  output_layer_names: "__kmax_seq_score_layer_0__"
   is_recurrent_layer_group: false
 }
 

python/paddle/trainer_config_helpers/tests/configs/protostr/test_pooling3D_layer.protostr

+type: "nn"
+layers {
+  name: "data_2d"
+  type: "data"
+  size: 6000
+  active_type: ""
+  height: 20
+  width: 10
+}
+layers {
+  name: "pool___2d"
+  type: "pool"
+  size: 840
+  active_type: ""
+  inputs {
+    input_layer_name: "data_2d"
+    pool_conf {
+      pool_type: "avg-projection"
+      channels: 30
+      size_x: 5
+      stride: 3
+      output_x: 4
+      img_size: 10
+      padding: 1
+      size_y: 5
+      stride_y: 3
+      output_y: 7
+      img_size_y: 20
+      padding_y: 1
+    }
+  }
+  height: 7
+  width: 4
+}
+layers {
+  name: "data_3d_1"
+  type: "data"
+  size: 60000
+  active_type: ""
+  height: 20
+  width: 10
+  depth: 10
+}
+layers {
+  name: "pool_3d_1"
+  type: "pool3d"
+  size: 3360
+  active_type: ""
+  inputs {
+    input_layer_name: "data_3d_1"
+    pool_conf {
+      pool_type: "avg-projection"
+      channels: 30
+      size_x: 5
+      stride: 3
+      output_x: 4
+      img_size: 10
+      padding: 1
+      size_y: 5
+      stride_y: 3
+      output_y: 7
+      img_size_y: 20
+      padding_y: 1
+      size_z: 5
+      stride_z: 3
+      output_z: 4
+      img_size_z: 10
+      padding_z: 1
+    }
+  }
+  height: 7
+  width: 4
+  depth: 4
+}
+layers {
+  name: "pool_3d_2"
+  type: "pool3d"
+  size: 3360
+  active_type: ""
+  inputs {
+    input_layer_name: "data_3d_1"
+    pool_conf {
+      pool_type: "max-projection"
+      channels: 30
+      size_x: 5
+      stride: 3
+      output_x: 4
+      img_size: 10
+      padding: 1
+      size_y: 5
+      stride_y: 3
+      output_y: 7
+      img_size_y: 20
+      padding_y: 1
+      size_z: 5
+      stride_z: 3
+      output_z: 4
+      img_size_z: 10
+      padding_z: 1
+    }
+  }
+  height: 7
+  width: 4
+  depth: 4
+}
+input_layer_names: "data_2d"
+output_layer_names: "pool___2d"
+output_layer_names: "pool_3d_1"
+output_layer_names: "pool_3d_2"
+sub_models {
+  name: "root"
+  layer_names: "data_2d"
+  layer_names: "pool___2d"
+  layer_names: "data_3d_1"
+  layer_names: "pool_3d_1"
+  layer_names: "pool_3d_2"
+  input_layer_names: "data_2d"
+  output_layer_names: "pool___2d"
+  output_layer_names: "pool_3d_1"
+  output_layer_names: "pool_3d_2"
+  is_recurrent_layer_group: false
+}
+

python/paddle/trainer_config_helpers/tests/configs/test_conv3d_layer.py

+from paddle.trainer_config_helpers import *
+
+settings(batch_size=1000, learning_rate=1e-5)
+
+num_channels = 3
+filter_size = 3
+filter_size_y = 3
+filter_size_z = 3
+stride = 2
+stride_y = 2
+stride_z = 2
+padding = 1
+padding_y = 1
+padding_z = 1
+groups = 1
+
+data = data_layer(
+    name='data', size=12096 * num_channels, height=48, width=42, depth=6)
+# first
+conv3d_1 = img_conv3d_layer(
+    input=data,
+    name='conv3d_1',
+    num_filters=16,
+    num_channels=num_channels,
+    filter_size=filter_size,
+    stride=stride,
+    padding=padding,
+    groups=groups,
+    bias_attr=True,
+    shared_biases=True,
+    trans=False,
+    layer_type="conv3d",
+    act=LinearActivation())
+# second
+conv3d_2 = img_conv3d_layer(
+    input=data,
+    name='conv3d_2',
+    num_filters=16,
+    num_channels=num_channels,
+    filter_size=[filter_size, filter_size_y, filter_size_z],
+    stride=[stride, stride_y, stride_z],
+    padding=[padding, padding_y, padding_z],
+    groups=groups,
+    bias_attr=True,
+    shared_biases=True,
+    trans=False,
+    layer_type="conv3d",
+    act=LinearActivation())
+outputs(conv3d_2)

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

+#!/usr/bin/env python
+#coding=utf-8
+
+from paddle.trainer_config_helpers import *
+beam_size = 5
+
+# the first beam expansion.
+sentence_states = data_layer(name="sentence_states", size=32)
+sentence_scores = data_layer(name="sentence_scores", size=1)
+topk_sentence_ids = kmax_seq_score_layer(
+    input=sentence_scores, beam_size=beam_size)
+
+# the second beam expansion.
+topk_sen = sub_nested_seq_layer(
+    input=sentence_states, selected_indices=topk_sentence_ids)
+start_pos_scores = fc_layer(input=topk_sen, size=1, act=LinearActivation())
+topk_start_pos_ids = kmax_seq_score_layer(
+    input=sentence_scores, beam_size=beam_size)
+
+# the final beam expansion.
+topk_start_spans = seq_slice_layer(
+    input=topk_sen, starts=topk_start_pos_ids, ends=None)
+end_pos_scores = fc_layer(
+    input=topk_start_spans, size=1, act=LinearActivation())
+topk_end_pos_ids = kmax_seq_score_layer(
+    input=end_pos_scores, beam_size=beam_size)
+
+# define the cost
+sentence_idx = data_layer(name="sentences_ids", size=1)
+start_idx = data_layer(name="start_ids", size=1)
+end_idx = data_layer(name="end_ids", size=1)
+cost = cross_entropy_over_beam(input=[
+    BeamInput(
+        candidate_scores=sentence_scores,
+        selected_candidates=topk_sentence_ids,
+        gold=sentence_idx), BeamInput(
+            candidate_scores=start_pos_scores,
+            selected_candidates=topk_start_pos_ids,
+            gold=start_idx), BeamInput(
+                candidate_scores=end_pos_scores,
+                selected_candidates=topk_end_pos_ids,
+                gold=end_idx)
+])
+
+outputs(cost)

python/paddle/trainer_config_helpers/tests/configs/test_deconv3d_layer.py

+from paddle.trainer_config_helpers import *
+
+settings(batch_size=1000, learning_rate=1e-5)
+
+num_channels = 3
+filter_size = 3
+filter_size_y = 3
+filter_size_z = 3
+stride = 2
+stride_y = 2
+stride_z = 2
+padding = 1
+padding_y = 1
+padding_z = 1
+groups = 1
+
+data = data_layer(
+    name='data', size=12096 * num_channels, height=48, width=42, depth=6)
+
+# first
+deconv3d_1 = img_conv3d_layer(
+    input=data,
+    name='deconv3d_1',
+    num_filters=16,
+    num_channels=num_channels,
+    filter_size=filter_size,
+    stride=stride,
+    padding=padding,
+    groups=groups,
+    bias_attr=True,
+    shared_biases=True,
+    trans=True,
+    layer_type="deconv3d",
+    act=LinearActivation())
+# second
+deconv3d_2 = img_conv3d_layer(
+    input=data,
+    name='deconv3d_2',
+    num_filters=16,
+    num_channels=num_channels,
+    filter_size=[filter_size, filter_size_y, filter_size_z],
+    stride=[stride, stride_y, stride_z],
+    padding=[padding, padding_y, padding_z],
+    groups=groups,
+    bias_attr=True,
+    shared_biases=True,
+    trans=True,
+    layer_type="deconv3d",
+    act=LinearActivation())
+outputs(deconv3d_2)

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

@@ -4,6 +4,6 @@ from paddle.trainer_config_helpers import *
 
 data = data_layer(name="input_seq", size=128)
 scores = fc_layer(input=data, size=1, act=ExpActivation())
-kmax_seq_id = kmax_sequence_score_layer(input=scores, beam_size=5)
+kmax_seq_id = kmax_seq_score_layer(input=scores, beam_size=5)
 
 outputs(kmax_seq_id)

python/paddle/trainer_config_helpers/tests/configs/test_pooling3D_layer.py

+from paddle.trainer_config_helpers import *
+
+settings(batch_size=100, learning_rate=1e-5)
+
+data_2d = data_layer(name='data_2d', size=6000, height=20, width=10)
+
+pool_2d = img_pool_layer(
+    name="pool___2d",
+    input=data_2d,
+    num_channels=30,
+    pool_size=5,
+    stride=3,
+    padding=1,
+    pool_type=AvgPooling())
+outputs(pool_2d)
+
+data_3d = data_layer(
+    name='data_3d_1', size=60000, depth=10, height=20, width=10)
+
+pool_3d_1 = img_pool3d_layer(
+    name="pool_3d_1",
+    input=data_3d,
+    num_channels=30,
+    pool_size=5,
+    stride=3,
+    padding=1,
+    pool_type=AvgPooling())
+outputs(pool_3d_1)
+
+pool_3d_2 = img_pool3d_layer(
+    name="pool_3d_2",
+    input=data_3d,
+    num_channels=30,
+    pool_size=[5, 5, 5],
+    stride=[3, 3, 3],
+    padding=[1, 1, 1],
+    pool_type=MaxPooling())
+outputs(pool_3d_2)

python/paddle/trainer_config_helpers/tests/layers_test.py

@@ -17,3 +17,4 @@ from paddle.trainer.config_parser import parse_config_and_serialize
 if __name__ == '__main__':
     parse_config_and_serialize(
         'trainer_config_helpers/tests/layers_test_config.py', '')
+# layers_test_config.py

python/paddle/v2/__init__.py

@@ -78,6 +78,8 @@ def init(**kwargs):
 
     if 'use_gpu' in kwargs:
         cp.g_command_config_args['use_gpu'] = kwargs['use_gpu']
+    if 'use_mkldnn' in kwargs:
+        cp.g_command_config_args['use_mkldnn'] = kwargs['use_mkldnn']
     assert 'parallel_nn' not in kwargs, ("currently 'parallel_nn' is not "
                                          "supported in v2 APIs.")
 

python/paddle/v2/inference.py

@@ -70,7 +70,7 @@ class Inference(object):
                 item = [each_result[each_field] for each_field in field]
                 yield item
 
-    def infer(self, input, field='value', **kwargs):
+    def infer(self, input, field='value', flatten_result=True, **kwargs):
         """
         Infer a data by model.
         :param input: input data batch. Should be python iterable object.
@@ -83,7 +83,10 @@ class Inference(object):
                 retv = [[] for i in xrange(len(result))]
             for i, item in enumerate(result):
                 retv[i].append(item)
-        retv = [numpy.concatenate(out) for out in retv]
+
+        if flatten_result:
+            retv = [numpy.concatenate(out) for out in retv]
+
         if len(retv) == 1:
             return retv[0]
         else: