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

Conflicts:
	paddle/pybind/pybind.cc

.travis.yml

@@ -4,7 +4,6 @@ cache:
     - $HOME/.ccache
     - $HOME/.cache/pip
     - $TRAVIS_BUILD_DIR/build/third_party
-    - $TRAVIS_BUILD_DIR/build_android/third_party
 sudo: required
 dist: trusty
 os:
@@ -12,7 +11,6 @@ os:
 env:
   - JOB=build_doc
   - JOB=check_style
-  - JOB=build_android
 addons:
   apt:
     packages:
@@ -23,7 +21,6 @@ addons:
       - python
       - python-pip
       - python2.7-dev
-      - python-numpy
       - python-wheel
       - libboost-dev
       - curl
@@ -37,8 +34,8 @@ before_install:
   - if [[ "$JOB" == "check_style" ]]; then sudo ln -s /usr/bin/clang-format-3.8 /usr/bin/clang-format; fi
   # Paddle is using protobuf 3.1 currently. Protobuf 3.2 breaks the compatibility. So we specify the python
   # protobuf version.
-  - pip install -r $TRAVIS_BUILD_DIR/python/requirements.txt
-  - pip install wheel sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit LinkChecker
+  - sudo pip install -r $TRAVIS_BUILD_DIR/python/requirements.txt
+  - sudo pip install wheel sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit LinkChecker
   - curl https://glide.sh/get | bash
   - eval "$(GIMME_GO_VERSION=1.8.3 gimme)"
   - go get -u github.com/alecthomas/gometalinter

CMakeLists.txt

@@ -65,8 +65,8 @@ if(NOT CMAKE_BUILD_TYPE)
 endif()
 
 if(ANDROID)
-    if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "21")
-        message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 21")
+    if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "16")
+        message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 16")
     endif()
 
     set(WITH_GPU OFF CACHE STRING

Dockerfile.android

@@ -4,9 +4,15 @@ MAINTAINER PaddlePaddle Authors <paddle-dev@baidu.com>
 ARG UBUNTU_MIRROR
 RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ubuntu.com/ubuntu#${UBUNTU_MIRROR}#g' /etc/apt/sources.list; fi'
 
+# ENV variables
+ARG ANDROID_ABI
+
+ENV ANDROID_ABI=${ANDROID_ABI:-"armeabi-v7a"}
+
 ENV HOME=/root \
     ANDROID_NDK_HOME=/opt/android-ndk-linux \
-    ANDROID_STANDALONE_TOOLCHAIN=/opt/android-toolchain-gcc
+    ANDROID_ARM_STANDALONE_TOOLCHAIN=/opt/arm-toolchain \
+    ANDROID_ARM64_STANDALONE_TOOLCHAIN=/opt/arm64-toolchain
 
 RUN apt-get update && \
     apt-get install -y \
@@ -15,12 +21,11 @@ RUN apt-get update && \
     apt-get clean -y
 
 # Install Go and glide
-RUN wget -O go.tgz https://storage.googleapis.com/golang/go1.8.1.linux-amd64.tar.gz && \
-    tar -C /usr/local -xzf go.tgz && \
+RUN wget -qO- go.tgz https://storage.googleapis.com/golang/go1.8.1.linux-amd64.tar.gz | \
+    tar -xz -C /usr/local && \
     mkdir /root/gopath && \
     mkdir /root/gopath/bin && \
-    mkdir /root/gopath/src && \
-    rm go.tgz
+    mkdir /root/gopath/src
 ENV GOROOT=/usr/local/go GOPATH=/root/gopath
 # should not be in the same line with GOROOT definition, otherwise docker build could not find GOROOT.
 ENV PATH=${PATH}:${GOROOT}/bin:${GOPATH}/bin
@@ -42,7 +47,8 @@ RUN mkdir /opt/android-ndk-tmp && \
     wget -q https://dl.google.com/android/repository/android-ndk-r14b-linux-x86_64.zip && \
     unzip -q android-ndk-r14b-linux-x86_64.zip && \
     mv android-ndk-r14b ${ANDROID_NDK_HOME} && \
-    ${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh --arch=arm --platform=android-21 --install-dir=${ANDROID_STANDALONE_TOOLCHAIN} && \
+    ${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh --arch=arm --platform=android-23 --install-dir=${ANDROID_ARM_STANDALONE_TOOLCHAIN} && \
+    ${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh --arch=arm64 --platform=android-23 --install-dir=${ANDROID_ARM64_STANDALONE_TOOLCHAIN} && \
     rm -rf /opt/android-ndk-tmp && \
     rm -rf ${ANDROID_NDK_HOME}
 

cmake/cross_compiling/android.cmake

@@ -20,6 +20,7 @@
 # The supported variables are listed belows:
 # 
 # ANDROID_STANDALONE_TOOLCHAIN
+# ANDROID_TOOLCHAIN
 # ANDROID_ABI
 # ANDROID_NATIVE_API_LEVEL
 # ANDROID_ARM_MODE
@@ -57,6 +58,10 @@ IF(NOT DEFINED CMAKE_SYSTEM_VERSION AND ANDROID_NATIVE_API_LEVEL)
     ENDIF()
 ENDIF()
 
+IF(NOT DEFINED ANDROID_TOOLCHAIN)
+    SET(ANDROID_TOOLCHAIN clang)
+ENDIF()
+
 IF(NOT DEFINED ANDROID_ABI)
     SET(ANDROID_ABI "armeabi-v7a")
 ENDIF()
@@ -82,6 +87,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
             "${CMAKE_VERSION}), when cross-compiling for Android.")
 
     IF(ANDROID_STANDALONE_TOOLCHAIN)
+        # Use standalone toolchain
         SET(CMAKE_SYSROOT "${ANDROID_STANDALONE_TOOLCHAIN}/sysroot")
 
         IF(NOT CMAKE_SYSTEM_VERSION)
@@ -96,26 +102,44 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
         ENDIF()
 
         # Toolchain
-        SET(ANDROID_TOOLCHAIN "gcc")
         SET(ANDROID_TOOLCHAIN_ROOT ${ANDROID_STANDALONE_TOOLCHAIN})
-        IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$")
-            SET(ANDROID_TOOLCHAIN_NAME arm-linux-androideabi)
-            IF(ANDROID_ABI STREQUAL "armeabi")
-                SET(CMAKE_SYSTEM_PROCESSOR armv5te)
-            ELSEIF(ANDROID_ABI STREQUAL "armeabi-v7a")
-                SET(CMAKE_SYSTEM_PROCESSOR armv7-a)
-            ENDIF()
-        ENDIF()
-        IF(ANDROID_ABI STREQUAL "arm64-v8a")
-            SET(ANDROID_TOOLCHAIN_NAME aarch64-linux-android)
-            SET(CMAKE_SYSTEM_PROCESSOR aarch64)
+    ELSE(ANDROID_NDK)
+        # TODO: use android ndk
+    ENDIF()
+
+    IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$")
+        SET(ANDROID_TOOLCHAIN_NAME arm-linux-androideabi)
+        IF(ANDROID_ABI STREQUAL "armeabi")
+            SET(CMAKE_SYSTEM_PROCESSOR armv5te)
+            SET(ANDROID_CLANG_TRIPLE armv5te-none-linux-androideabi)
+        ELSEIF(ANDROID_ABI STREQUAL "armeabi-v7a")
+            SET(CMAKE_SYSTEM_PROCESSOR armv7-a)
+            SET(ANDROID_CLANG_TRIPLE armv7-none-linux-androideabi)
         ENDIF()
-        SET(ANDROID_TOOLCHAIN_PREFIX "${ANDROID_TOOLCHAIN_ROOT}/bin/${ANDROID_TOOLCHAIN_NAME}-")
+    ELSEIF(ANDROID_ABI STREQUAL "arm64-v8a")
+        SET(ANDROID_TOOLCHAIN_NAME aarch64-linux-android)
+        SET(CMAKE_SYSTEM_PROCESSOR aarch64)
+        SET(ANDROID_CLANG_TRIPLE aarch64-none-linux-android)
+    ELSE()
+        MESSAGE(FATAL_ERROR "Invalid Android ABI: ${ANDROID_ABI}.")
+    ENDIF()
+    SET(ANDROID_TOOLCHAIN_PREFIX "${ANDROID_TOOLCHAIN_ROOT}/bin/${ANDROID_TOOLCHAIN_NAME}-")
+
+    IF(ANDROID_TOOLCHAIN STREQUAL clang)
+        SET(ANDROID_C_COMPILER_NAME clang)
+        SET(ANDROID_CXX_COMPILER_NAME clang++)
+        SET(CMAKE_C_COMPILER_TARGET   ${ANDROID_CLANG_TRIPLE})
+        SET(CMAKE_CXX_COMPILER_TARGET ${ANDROID_CLANG_TRIPLE})
+    ELSEIF(ANDROID_TOOLCHAIN STREQUAL gcc)
+        SET(ANDROID_C_COMPILER_NAME gcc)
+        SET(ANDROID_CXX_COMPILER_NAME g++)
+    ELSE()
+        MESSAGE(FATAL_ERROR "Invalid Android toolchain: ${ANDROID_TOOLCHAIN}")
     ENDIF()
 
     # C compiler
     IF(NOT CMAKE_C_COMPILER)
-        SET(ANDROID_C_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}gcc")
+        SET(ANDROID_C_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}${ANDROID_C_COMPILER_NAME}")
     ELSE()
         GET_FILENAME_COMPONENT(ANDROID_C_COMPILER ${CMAKE_C_COMPILER} PROGRAM)
     ENDIF()
@@ -125,7 +149,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
 
     # CXX compiler
     IF(NOT CMAKE_CXX_COMPILER)
-        SET(ANDROID_CXX_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}g++")
+        SET(ANDROID_CXX_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}${ANDROID_CXX_COMPILER_NAME}")
     ELSE()
         GET_FILENAME_COMPONENT(ANDROID_CXX_COMPILER ${CMAKE_CXX_COMPILER} PROGRAM)
     ENDIF()
@@ -137,7 +161,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
     SET(CMAKE_CXX_COMPILER ${ANDROID_CXX_COMPILER} CACHE PATH "CXX compiler" FORCE)
 
     # Toolchain and ABI specific flags.
-    SET(ANDROID_COMPILER_FLAGS "-ffunction-sections -fdata-sections -finline-limit=64")
+    SET(ANDROID_COMPILER_FLAGS "-ffunction-sections -fdata-sections")
     SET(ANDROID_LINKER_FLAGS "-Wl,--gc-sections")
 
     IF(ANDROID_ABI STREQUAL "armeabi")
@@ -145,8 +169,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
              -march=armv5te
              -mtune=xscale
              -msoft-float)
-    ENDIF()
-    IF(ANDROID_ABI STREQUAL "armeabi-v7a")
+    ELSEIF(ANDROID_ABI STREQUAL "armeabi-v7a")
         LIST(APPEND ANDROID_COMPILER_FLAGS
              -march=armv7-a
              -mfloat-abi=softfp)
@@ -156,6 +179,8 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
             LIST(APPEND ANDROID_COMPILER_FLAGS -mfpu=vfpv3-d16)
         ENDIF()
         LIST(APPEND ANDROID_LINKER_FLAGS -Wl,--fix-cortex-a8)
+    ELSEIF(ANDROID_ABI STREQUAL "arm64-v8a")
+        LIST(APPEND ANDROID_COMPILER_FLAGS -march=armv8-a)
     ENDIF()
 
     IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$")
@@ -164,10 +189,18 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
         ELSE()
             LIST(APPEND ANDROID_COMPILER_FLAGS -mthumb)
         ENDIF()
+        IF(ANDROID_TOOLCHAIN STREQUAL clang)
+            # Disable integrated-as for better compatibility.
+            LIST(APPEND ANDROID_COMPILER_FLAGS -fno-integrated-as)
+        ENDIF()
     ENDIF()
 
-    IF(ANDROID_ABI STREQUAL "arm64-v8a")
-        LIST(APPEND ANDROID_COMPILER_FLAGS -march=armv8-a)
+    IF(ANDROID_TOOLCHAIN STREQUAL clang)
+        # CMake automatically forwards all compiler flags to the linker,
+        # and clang doesn't like having -Wa flags being used for linking.
+        # To prevent CMake from doing this would require meddling with
+        # the CMAKE_<LANG>_COMPILE_OBJECT rules, which would get quite messy.
+        LIST(APPEND ANDROID_LINKER_FLAGS -Qunused-arguments)
     ENDIF()
 
     STRING(REPLACE ";" " " ANDROID_COMPILER_FLAGS "${ANDROID_COMPILER_FLAGS}")

cmake/external/openblas.cmake

@@ -12,6 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+IF(USE_EIGEN_FOR_BLAS)
+    return()
+ENDIF(USE_EIGEN_FOR_BLAS)
+
 INCLUDE(cblas)
 
 IF(NOT ${CBLAS_FOUND})

doc/design/functions_operators_layers.md

@@ -86,12 +86,13 @@ def layer.fc(X):
 
 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:
 

doc/design/graph.md

-# Design Doc: Computations as Graphs
+# Design Doc: Computations as a Graph
 
 A primary goal of the refactorization of PaddlePaddle is a more flexible representation of deep learning computation, in particular, a graph of operators and variables, instead of sequences of layers as before.
 
@@ -8,6 +8,8 @@ This document explains that the construction of a graph as three steps:
 - construct the backward part
 - construct the optimization part
 
+## The Construction of a Graph
+
 Let us take the problem of image classification as a simple example.  The application program that trains the model looks like:
 
 ```python
@@ -25,7 +27,9 @@ The first four lines of above program build the forward part of the graph.
 
 ![](images/graph_construction_example_forward_only.png)
 
-In particular, the first line `x = layer.data("images")` creates variable x and a Feed operator that copies a column from the minibatch to x.  `y = layer.fc(x)` creates not only the FC operator and output variable y, but also two parameters, W and b.
+In particular, the first line `x = layer.data("images")` creates variable x and a Feed operator that copies a column from the minibatch to x.  `y = layer.fc(x)` creates not only the FC operator and output variable y, but also two parameters, W and b, and the initialization operators.
+
+Initialization operators are kind of "run-once" operators -- the `Run` method increments a class data member counter so to run at most once.  By doing so, a parameter wouldn't be initialized repeatedly, say, in every minibatch.
 
 In this example, all operators are created as `OpDesc` protobuf messages, and all variables are `VarDesc`.  These protobuf messages are saved in a `BlockDesc` protobuf message.
 
@@ -49,3 +53,18 @@ According to the chain rule of gradient computation, `ConstructBackwardGraph` wo
 For each parameter, like W and b created by `layer.fc`, marked as double circles in above graphs, `ConstructOptimizationGraph` creates an optimization operator to apply its gradient.  Here results in the complete graph:
 
 ![](images/graph_construction_example_all.png)
+
+## Block and Graph
+
+The word block and graph are interchangable in the desgin of PaddlePaddle.  A [Block[(https://github.com/PaddlePaddle/Paddle/pull/3708) is a metaphore of the code and local variables in a pair of curly braces in programming languages, where operators are like statements or instructions.  A graph of operators and variables is a representation of the block.
+
+A Block keeps operators in an array `BlockDesc::ops`
+
+```protobuf
+message BlockDesc {
+  repeated OpDesc ops = 1;
+  repeated VarDesc vars = 2;
+}
+```
+
+in the order that there appear in user programs, like the Python program at the beginning of this article.  We can imagine that in `ops`,  we have some forward operators, followed by some gradient operators, and then some optimization operators.

doc/design/images/graph_construction_example.dot

@@ -2,6 +2,8 @@ digraph ImageClassificationGraph {
         ///////// The forward part /////////
         FeedX [label="Feed", color=blue, shape=box];
         FeedY [label="Feed", color=blue, shape=box];
+        InitW [label="Init", color=blue, shape=diamond];
+        Initb [label="Init", color=blue, shape=diamond];
         FC [label="FC", color=blue, shape=box];
         MSE [label="MSE", color=blue, shape=box];
 
@@ -14,6 +16,8 @@ digraph ImageClassificationGraph {
 
         FeedX -> x -> FC -> y -> MSE -> cost [color=blue];
         FeedY -> l [color=blue];
+        InitW -> W [color=blue];
+        Initb -> b [color=blue];
         W -> FC [color=blue];
         b -> FC [color=blue];
         l -> MSE [color=blue];

doc/design/ops/dist_train.md

+# Design Doc: Operation Graph Based Parameter Server
+
+## Abstract
+
+We propose an approach to implement the parameter server. In this
+approach, there is no fundamental difference between the trainer and
+the parameter server: they both run subgraphs, but subgraphs of
+different purposes.
+
+## Background
+
+The previous implementations of the parameter server does not run a
+subgraph. parameter initialization, optimizer computation, network
+communication and checkpointing are implemented twice on both the
+trainer and the parameter server.
+
+It would be great if we can write code once and use them on both the
+trainer and the parameter server: reduces code duplication and
+improves extensibility. Given that after the current refactor, we are
+representing everything as a computing graph on the
+trainer. Representing everything as a computing graph on the parameter
+server becomes a natural extension.
+
+## Design
+
+### Graph Converter
+
+The *graph converter* converts the user-defined operation (OP) graph
+into subgraphs to be scheduled on different nodes with the following
+steps:
+
+1. OP placement: the OPs will be placed on different nodes according
+   to heuristic that minimizes estimated total computation
+   time. Currently we will use a simple heuristic that puts parameter
+   varable on parameter server workers and everything else on trainer
+   workers.
+
+1. Add communication OPs to enable the communication between nodes.
+
+We will need these OPs: *Send*, *Recv*, *Enqueue*, *Dequeue*.
+
+Below is an example of converting the user defined graph to the
+subgraphs for the trainer and the parameter server:
+
+<img src="src/local-graph.png" width="300"/>
+
+After converting:
+
+<img src="src/dist-graph.png" width="700"/>
+
+1. The parameter variable W and it's optimizer subgraph are placed on the parameter server.
+1. Operators are added to the subgraphs.
+   - *Send* sends data to the connected *Recv* operator.  The
+	 scheduler on the receive node will only schedule *Recv* operator
+	 to run when the *Send* operator has ran (the *Send* OP will mark
+	 the *Recv* OP runnable automatically).
+   - *Enueue* enqueues the input variable, it can block until space
+     become available in the queue.
+   - *Dequeue* outputs configurable numbers of tensors from the
+     queue. It will block until the queue have the required number of
+     tensors.
+
+
+### Benefits
+
+- Model parallelism become easier to implement: it's an extension to
+  the trainer - parameter server approach. we already have the
+  communication OPs, but need to extend the graph converter's
+  placement functionality.
+
+- User-defined optimizer is easier to add - user can now express it as
+  a subgraph.
+
+- No more duplication logic inside the trainer and the parameter
+  server mentioned in the background section.
+
+### Challenges
+
+- It might be hard for the graph converter to cut a general graph
+  (without any hint for which subgraph is the optimizer). We may need
+  to label which subgraph inside the OP graph is the optimizer.
+
+- It's important to balance the parameter shards of on multiple
+  parameter server. If a single parameter is very big (some
+  word-embedding, fully connected, softmax layer), we need to
+  automatically partition the single parameter onto different
+  parameter servers when possible (only element-wise optimizer depends
+  on the parameter variable).
+
+### Discussion
+
+- In the "Aync SGD" figure, the "W" variable on the parameter server
+  could be read and wrote concurrently, what is our locking strategy?
+  E.g., each variable have a lock cpp method to be invoked by every
+  OP, or, have a lock OP.
+
+- Can the Enqueue OP be implemented under our current tensor design
+  (puts the input tensor into the queue tensor)?
+
+- *Dequeue* OP will have variable numbers of output (depends on the
+  `min_count` attribute), does our current design support it? (similar
+  question for the *Add* OP)
+
+
+### References:
+[1] [TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems](https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/45166.pdf)

doc/design/simple_op_design.md

@@ -147,7 +147,7 @@ class CosineOp {
 struct CosineOpProtoMaker : public OpProtoMaker {
 	CosineOpProtoMaker(OpProto* proto) : OpProtoMaker(proto) {
 		AddInput("input", "input of cosine op");
-		AddAttr("scale", "scale of cosine op", float).Default(1.0).LargerThan(0.0);
+		AddAttr("scale", "scale of cosine op", float).Default(1.0).GreaterThan(0.0);
 		AddType("cos");
 		AddComment("This is cos op");
 	}

doc/design/var_desc.md

+## Background
+PaddlePaddle divides the description of neural network computation graph into two stages: compile time and runtime.
+
+PaddlePaddle use proto message to describe compile time graph for
+
+1. Computation graph should be able to be saved to a file.
+1. In distributed training, the graph will be serialized and send to multiple workers.
+
+The computation graph is constructed by Data Node and Operation Node. The concept to represent them is in the table below.
+
+| |compile time|runtime|
+|---|---|---|
+|Data|VarDesc(proto)|Variable(cpp)|
+|Operation|OpDesc(proto)|Operator(cpp)|
+
+
+## Definition of VarDesc
+
+A VarDesc should have a name and value, in PaddlePaddle, the value will always be a tensor. Since we use LoDTensor most of the time. We add a LoDTesnorDesc to represent it.
+
+```proto
+message VarDesc {
+  required string name = 1;
+  optional LoDTensorDesc lod_tensor = 2;
+}
+```
+
+## Definition of LodTensorDesc
+
+```proto
+enum DataType {
+  BOOL = 0;
+  INT16 = 1;
+  INT32 = 2;
+  INT64 = 3;
+  FP16 = 4;
+  FP32 = 5;
+  FP64 = 6;
+}
+
+message LoDTensorDesc {
+  required DataType data_type = 1;
+  repeated int32 dims = 2; // [UNK, 640, 480] is saved as [-1, 640, 480]
+  optional int32 lod_level = 3 [default=0];
+}
+```
+
+## Definition of Variable in Python
+
+In Python API, layer will take Variable as Input, and return Variable as Output. There should be a class `Variable` in python to help create and manage Variable.
+
+```python
+image = Variable(dims=[-1, 640, 480])
+# fc1 and fc2 are both Variable
+fc1 = layer.fc(input=image, output_size=10)
+fc2 = layer.fc(input=fc1, output_size=20)
+```
+### what should class `Variable` Have
+1. `name`.a name of string type is used to mark the value of the Variable.
+1. `initializer`. Since our Tensor does not have value. we will always use some Operator to fullfill it when run. So we should have a initialize method to help add the init operator.
+1. `operator`. Variable should record which operator produce itself. The reaon is:
+  - we use pd.eval(targets=[var1, var2]) to run the related ops to get the value of var1 and var2. var.op is used to trace the dependency of the current variable.
+
+In PaddlePaddle, we use Block to describe Computation Graph, so in the code we will use Block but not Graph.
+
+```python
+import VarDesc
+import LoDTensorDesc
+import framework
+
+def AddInitialOperator(variable, initializer):
+	# add an initialize Operator to block to init this Variable
+
+class Variable(object):
+   def __init__(self, name, dims, type, initializer):
+      self._block = get_default_block()
+      self._name = name
+      self.op = None
+
+      tensor_desc = LoDTensorDesc(data_type=type, dims=dims)
+      _var_desc = VarDesc(name=name, lod_tensor=tensor_desc)
+      self._var = framework.CreateVar(_var_desc)
+      self._block.add_var(self)
+
+      # add initial op according to initializer
+      if initializer is not None:
+          AddInitialOperator(self, initializer)
+
+   def dims(self):
+      return self._var.dims()
+
+   def data_type(self):
+       return self._var.data_type()
+
+   def to_proto(self):
+       pass
+```
+
+Then we can use this Variable to create a fc layer in Python.
+
+```python
+import paddle as pd
+
+def flatten_size(X, num_flatten_dims):
+  prod = 1 # of last num_flatten_dims
+  for i in xrange(num_flatten_dims):
+    prod = prod * X.dims[-i-1]
+  return prod
+
+def layer.fc(X, output_size, num_flatten_dims):
+  W = Variable(pd.random_uniform(), type=FP32, dims=[flatten_size(X, num_flatten_dims), output_size])
+  b = Variable(pd.random_uniform(), type=FP32, dims=[output_size])
+  out = Variable(type=FP32)
+  y = operator.fc(X, W, b, output=out) # fc will put fc op input into out
+  pd.InferShape(y)
+  return out
+
+x = Variable(dims=[-1, 640, 480])
+y = layer.fc(x, output_size=100)
+z = layer.fc(y, output_size=200)
+
+paddle.eval(targets=[z], ...)
+print(z)
+```

paddle/cuda/include/hl_cpu_gru.cuh

@@ -18,14 +18,6 @@ limitations under the License. */
 
 #ifndef __NVCC__
 
-#include "paddle/math/MathFunctions.h"
-
-#ifndef PADDLE_TYPE_DOUBLE
-#define     CBLAS_GEMM     paddle::gemm<float>
-#else
-#define     CBLAS_GEMM     paddle::gemm<double>
-#endif
-
 template<class OpResetOutput>
 void hl_naive_gru_forward_reset_output(OpResetOutput opResetOutput,
                                        real *gateValue,
@@ -210,51 +202,6 @@ inline void forward_final_output(OpFinalOutput opFinalOutput,
   }
 }
 
-template<class OpResetOutput, class OpFinalOutput>
-void hl_cpu_gru_forward(OpResetOutput opResetOutput,
-                        OpFinalOutput opFinalOutput,
-                        hl_gru_value value,
-                        int frameSize,
-                        int batchSize,
-                        hl_activation_mode_t active_node,
-                        hl_activation_mode_t active_gate) {
-  if (value.prevOutValue) {
-    CBLAS_GEMM(CblasNoTrans,
-               CblasNoTrans,
-               batchSize,
-               2 * frameSize,
-               frameSize,
-               1,
-               value.prevOutValue,
-               frameSize,
-               value.gateWeight,
-               frameSize * 2,
-               1,
-               value.gateValue,
-               frameSize * 3);
-  }
-
-  forward_reset_output(opResetOutput, value, frameSize, batchSize, active_gate);
-
-  if (value.prevOutValue) {
-    CBLAS_GEMM(CblasNoTrans,
-               CblasNoTrans,
-               batchSize,
-               frameSize,
-               frameSize,
-               1,
-               value.resetOutputValue,
-               frameSize,
-               value.stateWeight,
-               frameSize,
-               1,
-               value.gateValue + frameSize * 2,
-               frameSize * 3);
-  }
-
-  forward_final_output(opFinalOutput, value, frameSize, batchSize, active_node);
-}
-
 template<class OpStateGrad>
 void hl_naive_gru_backward_state_grad(OpStateGrad opStateGrad,
                                       real *gateValue,
@@ -525,86 +472,6 @@ inline void backward_reset_grad(OpResetGrad opResetGrad,
   }
 }
 
-template<class OpStateGrad, class OpResetGrad>
-void hl_cpu_gru_backward(OpStateGrad opStateGrad,
-                         OpResetGrad opResetGrad,
-                         hl_gru_value value,
-                         hl_gru_grad  grad,
-                         int frameSize,
-                         int batchSize,
-                         hl_activation_mode_t active_node,
-                         hl_activation_mode_t active_gate) {
-  backward_state_grad(opStateGrad, value, grad,
-    frameSize, batchSize, active_node);
-
-  if (value.prevOutValue && grad.prevOutGrad) {
-    CBLAS_GEMM(CblasNoTrans,
-               CblasTrans,
-               batchSize,
-               frameSize,
-               frameSize,
-               1,
-               grad.gateGrad + frameSize * 2,
-               frameSize * 3,
-               value.stateWeight,
-               frameSize,
-               0,
-               grad.resetOutputGrad,
-               frameSize);
-
-    if (grad.stateWeightGrad) {
-      CBLAS_GEMM(CblasTrans,
-                 CblasNoTrans,
-                 frameSize,
-                 frameSize,
-                 batchSize,
-                 1,
-                 value.resetOutputValue,
-                 frameSize,
-                 grad.gateGrad + frameSize * 2,
-                 frameSize * 3,
-                 1,
-                 grad.stateWeightGrad,
-                 frameSize);
-    }
-  }
-
-  backward_reset_grad(opResetGrad, value, grad,
-    frameSize, batchSize, active_gate);
-
-  if (grad.prevOutGrad && value.prevOutValue) {
-    CBLAS_GEMM(CblasNoTrans,
-               CblasTrans,
-               batchSize,
-               frameSize,
-               frameSize * 2,
-               1,
-               grad.gateGrad,
-               frameSize * 3,
-               value.gateWeight,
-               frameSize * 2,
-               1,
-               grad.prevOutGrad,
-               frameSize);
-
-    if (grad.gateWeightGrad) {
-      CBLAS_GEMM(CblasTrans,
-                 CblasNoTrans,
-                 frameSize,
-                 frameSize * 2,
-                 batchSize,
-                 1,
-                 value.prevOutValue,
-                 frameSize,
-                 grad.gateGrad,
-                 frameSize * 3,
-                 1,
-                 grad.gateWeightGrad,
-                 frameSize * 2);
-    }
-  }
-}
-
 #endif
 
 #endif  // HL_CPU_GRU_CUH_

paddle/framework/attribute.h

@@ -41,11 +41,23 @@ Attribute GetAttrValue(const OpDesc::Attr& attr_desc);
 
 // check whether a value(attribute) fit a certain limit
 template <typename T>
-class LargerThanChecker {
+class GreaterThanChecker {
  public:
-  explicit LargerThanChecker(T lower_bound) : lower_bound_(lower_bound) {}
+  explicit GreaterThanChecker(T lower_bound) : lower_bound_(lower_bound) {}
   void operator()(T& value) const {
-    PADDLE_ENFORCE(value > lower_bound_, "larger_than check fail");
+    PADDLE_ENFORCE(value > lower_bound_, "larger_than check fails.");
+  }
+
+ private:
+  T lower_bound_;
+};
+
+template <typename T>
+class EqualGreaterThanChecker {
+ public:
+  explicit EqualGreaterThanChecker(T lower_bound) : lower_bound_(lower_bound) {}
+  void operator()(T& value) const {
+    PADDLE_ENFORCE_GE(value, lower_bound_, "equal_larger_than check fails.");
   }
 
  private:
@@ -110,8 +122,13 @@ class TypedAttrChecker {
     return *this;
   }
 
-  TypedAttrChecker& LargerThan(const T& lower_bound) {
-    value_checkers_.push_back(LargerThanChecker<T>(lower_bound));
+  TypedAttrChecker& GreaterThan(const T& lower_bound) {
+    value_checkers_.push_back(GreaterThanChecker<T>(lower_bound));
+    return *this;
+  }
+
+  TypedAttrChecker& EqualGreaterThan(const T& lower_bound) {
+    value_checkers_.push_back(EqualGreaterThanChecker<T>(lower_bound));
     return *this;
   }
 

paddle/framework/backward.md

@@ -2,20 +2,20 @@
 
 ## Motivation
 
-In Neural Network, the backpropagation algorithm follows the chain rule, so we need to compound the fundmental gradient operators/expressions together with chain rule . Every forward network need a backward network to construct the full computation graph, the operator/expression's backward pass will be generated respect to forward pass.
-  
+In Neural Network, the backpropagation algorithm follows the chain rule, so we need to compound the gradient operators/expressions together with the chain rule. Every forward network needs a backward network to construct the full computation graph, the operator/expression's backward pass will be generated respect to forward pass.
+
 ## Backward Operator Registry
 
-A backward network is built up with several backward operators. Backward operators take forward operators' inputs, outputs and output gradients and then calculate its input gradients.
+A backward network is built up with several backward operators. Backward operators take forward operators' inputs outputs, and output gradients and then calculate its input gradients.
 
 |                        | forward operator | backward operator 
 | ---------------------- | ---------------- |------------------------- |		
 | **Operator::inputs_**  | Inputs       | Inputs, Outputs, OutputGradients |	
 | **Operator::outputs_** | Outputs          | InputGradients            |
 
- In most cases, there is a one-to-one correspondence between forward and backward operators. These correspondences are recorded by a global hash map(`OpInfoMap`). To follow the philosophy of minimum core and make operators pluggable, the registry mechanism is introduced.
+ In most cases, there is a one-to-one correspondence between the forward and backward operators. These correspondences are recorded by a global hash map(`OpInfoMap`). To follow the philosophy of minimum core and make operators pluggable, the registry mechanism is introduced.
 
-For example, we have got a `mul_op`, and we can register it's information and corresponding backward operator by the following macro:
+For example, we have got a `mul_op`, and we can register its information and corresponding backward operator by the following macro:
 
 ```cpp
 REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad);
@@ -27,17 +27,17 @@ REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad);
 
 ## Backward Opeartor Creating
 
-Given a certain forward operator, we can get its corresponding backward opeartor by calling:
+Given a certain forward operator, we can get its corresponding backward operator by calling:
 
 ```cpp
 OperatorBase* bwd_op = BuildGradOp(const OperatorBase* fwd_op);
-``` 
+```
 
 The function `BuildGradOp` will sequentially execute following processes:
 
 1. Get the `type_` of given forward operator, and then get the corresponding backward operator's type by looking up the `OpInfoMap`.
 
-2. Build two maps named `inputs` and `outputs` to temporary storage backward operator's inputs and outputs. Copy forward operator's `inputs_` and `outputs_` to map `inputs`, except these are not necessary for gradient computing.
+2. Build two maps named `inputs` and `outputs` to temporary storage backward operator's inputs and outputs. Copy forward operator's `inputs_` and `outputs_` to map `inputs`, except these, are not necessary for gradient computing.
 
 3. Add forward inputs' gradient variables into map `output`, adding forward outputs' gradient variables into map `input`.
 
@@ -49,31 +49,31 @@ A backward network is a series of backward operators. The main idea of building
 
 In our design, the network itself is also a kind of operator. So the operators contained by a big network may be some small network. 
 
-given a forward network, it generates the backward network. We only care about the Gradients—`OutputGradients`,`InputGradients`.
+given a forward network, it generates the backward network. We only care about the Gradients—`OutputGradients`, `InputGradients`.
 
 1. Op 
 
-   when the input forward network is a Op, return its gradient Operator Immediately.
+   when the input forward network is an Op, return its gradient Operator Immediately.
 
 2. NetOp 
 
-   when the input forward network is a NetOp, it need to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to forward NetOp.
+   when the input forward network is a NetOp, it needs to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to the forward NetOp.
 
-   **shared variable**. As illustrated in the pictures, two operator's `Output` `Gradient` will overwirte their shared input variable.  
+   **shared variable**. As illustrated in the pictures, two operator's `Output` `Gradient` will overwrite their shared input variable.  
 
    <p align="center">
-   <img src="./images/duplicate_op.png" width="70%" ><br/>
+   <img src="./images/duplicate_op.png" width="50%" ><br/>
 
-   1. shared variable in two operators. 
+   1. Shared variable in operators. 
 
    </p>
 
-   Share variable between operators or same input variable used in multiple operators lead to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively, and add a generic add operator replace the overwirte links. 
+   Share variable between operators or same input variable used in multiple operators leads to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively and add a generic add operator replace the overwrite links. 
 
    <p align="center">
-   <img src="images/duplicate_op2.png" width="90%" ><br/>
+   <img src="images/duplicate_op2.png" width="50%" ><br/>
 
-   2. replace shared variable gradient with `Add` Operator
+   2. Replace shared variable's gradient with `Add` operator.
 
    </p>
 

paddle/framework/ddim.cc

@@ -283,5 +283,14 @@ std::ostream& operator<<(std::ostream& os, const DDim& ddim) {
 DDim::DDim(std::initializer_list<int64_t> init_list) {
   *this = make_ddim(init_list);
 }
+
+DDim flatten_to_2d(const DDim& src, int num_col_dims) {
+  int rank = src.size();
+  return make_ddim({product(slice_ddim(src, 0, num_col_dims)),
+                    product(slice_ddim(src, num_col_dims, rank))});
+}
+
+DDim flatten_to_1d(const DDim& src) { return make_ddim({product(src)}); }
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/ddim.h

@@ -115,6 +115,12 @@ int arity(const DDim& ddim);
 
 std::ostream& operator<<(std::ostream&, const DDim&);
 
+// Reshape a tensor to a matrix. The matrix's first dimension(column length)
+// will be the product of tensor's first `num_col_dims` dimensions.
+DDim flatten_to_2d(const DDim& src, int num_col_dims);
+
+DDim flatten_to_1d(const DDim& src);
+
 }  // namespace framework
 }  // namespace paddle
 

paddle/framework/eigen.h

@@ -63,20 +63,35 @@ struct EigenTensor {
 
 template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
-struct EigenMatrix : public EigenTensor<T, 2, MajorType, IndexType> {};
+struct EigenMatrix : public EigenTensor<T, 2, MajorType, IndexType> {
+  static typename EigenMatrix::Type Reshape(Tensor& tensor, int num_col_dims) {
+    int rank = tensor.dims_.size();
+    PADDLE_ENFORCE(num_col_dims > 0 && num_col_dims < rank,
+                   "`num_col_dims` must be between (0, rank_of_tensor).");
+    return EigenMatrix::From(tensor,
+                             flatten_to_2d(tensor.dims(), num_col_dims));
+  }
+
+  static typename EigenMatrix::ConstType Reshape(const Tensor& tensor,
+                                                 int num_col_dims) {
+    int rank = tensor.dims_.size();
+    PADDLE_ENFORCE(num_col_dims > 0 && num_col_dims < rank,
+                   "`num_col_dims` must be between (0, rank_of_tensor).");
+    return EigenMatrix::From(tensor,
+                             flatten_to_2d(tensor.dims(), num_col_dims));
+  }
+};
 
 template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 struct EigenVector : public EigenTensor<T, 1, MajorType, IndexType> {
   // Flatten reshapes a Tensor into an EigenVector.
   static typename EigenVector::Type Flatten(Tensor& tensor) {
-    return EigenVector::From(
-        tensor, make_ddim({static_cast<int>(product(tensor.dims_))}));
+    return EigenVector::From(tensor, {product(tensor.dims_)});
   }
 
   static typename EigenVector::ConstType Flatten(const Tensor& tensor) {
-    return EigenVector::From(
-        tensor, make_ddim({static_cast<int>(product(tensor.dims_))}));
+    return EigenVector::From(tensor, {product(tensor.dims_)});
   }
 };
 

paddle/framework/eigen_test.cc

@@ -108,5 +108,24 @@ TEST(Eigen, Matrix) {
   }
 }
 
+TEST(Eigen, MatrixReshape) {
+  Tensor t;
+  float* p = t.mutable_data<float>({2, 3, 6, 4}, platform::CPUPlace());
+  for (int i = 0; i < 2 * 3 * 6 * 4; ++i) {
+    p[i] = static_cast<float>(i);
+  }
+
+  EigenMatrix<float>::Type em = EigenMatrix<float>::Reshape(t, 2);
+
+  ASSERT_EQ(2 * 3, em.dimension(0));
+  ASSERT_EQ(6 * 4, em.dimension(1));
+
+  for (int i = 0; i < 2 * 3; i++) {
+    for (int j = 0; j < 6 * 4; j++) {
+      ASSERT_NEAR(i * 6 * 4 + j, em(i, j), 1e-6f);
+    }
+  }
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/framework.proto

@@ -87,3 +87,24 @@ message OpProto {
   repeated Attr attrs = 4;
   required string comment = 5;
 }
+
+enum DataType {
+  BOOL = 0;
+  INT16 = 1;
+  INT32 = 2;
+  INT64 = 3;
+  FP16 = 4;
+  FP32 = 5;
+  FP64 = 6;
+}
+
+message LoDTensorDesc {
+  required DataType data_type = 1;
+  repeated int32 dims = 2; // [UNK, 640, 480] is saved as [-1, 640, 480]
+  optional int32 lod_level = 3 [ default = 0 ];
+}
+
+message VarDesc {
+  required string name = 1;
+  optional LoDTensorDesc lod_tensor = 2;
+}

paddle/framework/grad_op_builder_test.cc

@@ -3,7 +3,7 @@
 #include "paddle/framework/op_registry.h"
 #include "paddle/framework/operator.h"
 
-USE_OP(add_two);
+USE_OP(add);
 
 namespace paddle {
 namespace framework {
@@ -41,7 +41,7 @@ namespace f = paddle::framework;
 
 TEST(GradOpBuilder, AddTwo) {
   std::shared_ptr<f::OperatorBase> add_op(f::OpRegistry::CreateOp(
-      "add_two", {{"X", {"x"}}, {"Y", {"y"}}}, {{"Out", {"out"}}}, {}));
+      "add", {{"X", {"x"}}, {"Y", {"y"}}}, {{"Out", {"out"}}}, {}));
   std::shared_ptr<f::OperatorBase> grad_add_op =
       f::OpRegistry::CreateGradOp(*add_op);
   EXPECT_EQ(grad_add_op->Inputs().size(), 4UL);

paddle/framework/lod_tensor.cc

@@ -19,8 +19,8 @@
 namespace paddle {
 namespace framework {
 
-LOD SliceLevels(const LOD& in, size_t level_begin, size_t level_end) {
-  LOD new_lod;
+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(in.at(i));
@@ -28,10 +28,10 @@ LOD SliceLevels(const LOD& in, size_t level_begin, size_t level_end) {
   return new_lod;
 }
 
-LOD SliceInLevel(const LOD& in, size_t level, size_t elem_begin,
+LoD SliceInLevel(const LoD& in, size_t level, size_t elem_begin,
                  size_t elem_end) {
   // slice the lod.
-  LOD new_lod;
+  LoD new_lod;
   new_lod.reserve(in.size() - level);
   auto start = in.at(level)[elem_begin];
   auto end = in.at(level)[elem_end];
@@ -46,13 +46,13 @@ LOD SliceInLevel(const LOD& in, size_t level, size_t elem_begin,
     std::transform(new_lod.back().begin(), new_lod.back().end(),
                    new_lod.back().begin(),
                    [start](int v) { return v - start; });
-    PADDLE_ENFORCE_EQ(new_lod.back().front(), 0, "error in slice LOD");
+    PADDLE_ENFORCE_EQ(new_lod.back().front(), 0, "error in slice LoD");
   }
   PADDLE_ENFORCE_LE(new_lod.size(), in.size());
   return new_lod;
 }
 
-bool operator==(const LOD& a, const LOD& b) {
+bool operator==(const LoD& a, const LoD& b) {
   if (a.size() != b.size()) {
     return false;
   }
@@ -72,12 +72,12 @@ bool operator==(const LOD& a, const LOD& b) {
   return true;
 }
 
-void LODTensor::SliceLevels(size_t level_begin, size_t level_end) {
+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) {
+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),

paddle/framework/lod_tensor.h

@@ -35,34 +35,34 @@ template <typename T>
 using Vector = thrust::host_vector<T>;
 #endif
 
-using LOD = std::vector<Vector<size_t>>;
+using LoD = std::vector<Vector<size_t>>;
 
-LOD SliceLevels(const LOD& in, size_t level_begin, size_t level_end);
+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,
+LoD SliceInLevel(const LoD& in, size_t level, size_t elem_begin,
                  size_t elem_end);
 
-bool operator==(const LOD& a, const LOD& b);
+bool operator==(const LoD& a, const LoD& b);
 
 /*
- * LODTensor (Level of details Tensor)
+ * LoDTensor (Level of details Tensor)
  * see https://en.wikipedia.org/wiki/Level_of_details for reference.
  */
-class LODTensor {
+class LoDTensor {
  public:
-  LODTensor() {}
-  LODTensor(const LOD& lod, Tensor* t) : lod_(lod), tensor_(t) {}
+  LoDTensor() {}
+  LoDTensor(const LoD& lod, Tensor* t) : lod_(lod), tensor_(t) {}
 
-  void set_lod(const LOD& lod) { lod_ = lod; }
+  void set_lod(const LoD& lod) { lod_ = lod; }
 
   void set_tensor(Tensor* tensor) { tensor_ = tensor; }
 
   Tensor& tensor() { return *tensor_; }
 
-  LOD lod() { return lod_; }
+  LoD lod() { return lod_; }
 
   /*
-   * Get a element from LOD.
+   * Get a element from LoD.
    */
   size_t lod_element(size_t level, size_t elem) const {
     PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
@@ -74,7 +74,7 @@ class LODTensor {
   }
 
   /*
-   * Number of LODTensor's levels, each level has units of data, for example,
+   * Number of LoDTensor's levels, each level has units of data, for example,
    * in the sentence's view, article, paragraph, sentence are 3 levels.
    */
   size_t NumLevels() const { return lod_.size(); }
@@ -100,7 +100,7 @@ class LODTensor {
   void SliceInLevel(size_t level, size_t elem_begin, size_t elem_end);
 
  private:
-  LOD lod_;
+  LoD lod_;
   Tensor* tensor_;  // not owned
 };
 }  // namespace framework

paddle/framework/lod_tensor.md

@@ -94,7 +94,7 @@ Let's go on slicing this slice.  Its <1,1>-slice is
 |||
 ```
 
-### The General Slicing Algorithm
+### The Slicing Algorithm
 
 The algorithm, with over-simplified data structure, is defined as
 
@@ -106,17 +106,41 @@ struct LoDTensor {
   float* tensor_;
 };
 
-LoDTensor Slice(const LoDTensor& lodt, int level, int sequence) {
+LoDTensor Slice(const LoDTensor& lodt, int level, int sequence);
+```
+
+Let us revisit the example above
 
-}
+```
+         3
+3           1  2
+3   2  4    1  2  3
+||| || |||| |  || |||
 ```
 
-### Slicing the Top Level
+Suppose that we want to retrieve the <1,2>-slice
 
-Please be aware that an RNN operator only slices the top level of a LoD Tensor to get the step inputs.
+```
+2
+2  3
+|| |||
+```
 
-```c++
-LoDTensor Slice(const LoDTensor& lodt, int sequence) {
+we will need to find out the starting position of this slice by summing over all leaf nodes in `LoD` to the left of the slice, i.e., 3 + 2 + 4 + 1 = 10.
+
+To avoid the traversal of the LoD tree at slcing time,  we can do it at the construction time -- instead of saving the lengths of the next level in the LoD tree, we can save the starting offset of the next level.  For example, above LoD Tensor can be transformed into
+
+```
+        0
+0           9  10
+0   3  5    9  10 12
+||| || |||| |  || |||
+```
+
+We don't really need the 0 on top, so the LoD Tensor could be
 
-}
+```
+0           9  10
+0   3  5    9  10 12
+||| || |||| |  || |||
 ```

paddle/framework/lod_tensor_test.cc

@@ -21,7 +21,7 @@
 namespace paddle {
 namespace framework {
 
-class LODTensorTester : public ::testing::Test {
+class LoDTensorTester : public ::testing::Test {
  public:
   virtual void SetUp() override {
     // tensor's batch_size: 30
@@ -29,7 +29,7 @@ class LODTensorTester : public ::testing::Test {
     // 0 10 20
     // 0 5 10 15 20
     // 0 2 5 7 10 12 15 20
-    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});
@@ -47,21 +47,21 @@ class LODTensorTester : public ::testing::Test {
  protected:
   platform::CPUPlace place;
   Tensor tensor;
-  LODTensor lod_tensor;
+  LoDTensor lod_tensor;
 };
 
-TEST_F(LODTensorTester, NumLevels) { ASSERT_EQ(lod_tensor.NumLevels(), 3UL); }
+TEST_F(LoDTensorTester, NumLevels) { ASSERT_EQ(lod_tensor.NumLevels(), 3UL); }
 
-TEST_F(LODTensorTester, NumElements) {
+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);
 }
 
-TEST_F(LODTensorTester, SliceLevels) {
+TEST_F(LoDTensorTester, SliceLevels) {
   // slice 1 level
   for (size_t level = 0; level < 3UL; ++level) {
-    LODTensor new_lod_tensor = lod_tensor;
+    LoDTensor new_lod_tensor = lod_tensor;
     new_lod_tensor.SliceLevels(level, level + 1);
     ASSERT_EQ(new_lod_tensor.NumLevels(), 1UL);
     ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor.NumElements(level));
@@ -70,7 +70,7 @@ TEST_F(LODTensorTester, SliceLevels) {
   }
   // slice 2 level
   for (size_t level = 0; level < 2UL; ++level) {
-    LODTensor new_lod_tensor = lod_tensor;
+    LoDTensor new_lod_tensor = lod_tensor;
     new_lod_tensor.SliceLevels(level, level + 2);
     ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
     ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor.NumElements(level));
@@ -80,9 +80,9 @@ TEST_F(LODTensorTester, SliceLevels) {
   }
 }
 
-TEST_F(LODTensorTester, SliceInLevel) {
+TEST_F(LoDTensorTester, SliceInLevel) {
   size_t level = 0;
-  LODTensor new_lod_tensor = lod_tensor;
+  LoDTensor new_lod_tensor = lod_tensor;
   new_lod_tensor.SliceInLevel(level, 0, 2);
   EXPECT_EQ(new_lod_tensor.NumLevels(), 3UL);
   EXPECT_EQ(new_lod_tensor.NumElements(0), 2UL);

paddle/framework/op_registry_test.cc

@@ -21,7 +21,7 @@ class CosineOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
     AddOutput("output", "output of cosine op");
     AddAttr<float>("scale", "scale of cosine op")
         .SetDefault(1.0)
-        .LargerThan(0.0);
+        .GreaterThan(0.0);
     AddComment("This is cos op");
   }
 };
@@ -80,7 +80,7 @@ TEST(OpRegistry, CreateOp) {
   paddle::framework::Scope scope;
   paddle::platform::CPUDeviceContext dev_ctx;
   op->Run(scope, dev_ctx);
-  float scale_get = op->GetAttr<float>("scale");
+  float scale_get = op->Attr<float>("scale");
   ASSERT_EQ(scale_get, scale);
 }
 
@@ -121,7 +121,7 @@ TEST(OpRegistry, DefaultValue) {
   paddle::framework::Scope scope;
   paddle::platform::CPUDeviceContext dev_ctx;
   op->Run(scope, dev_ctx);
-  ASSERT_EQ(op->GetAttr<float>("scale"), 1.0);
+  ASSERT_EQ(op->Attr<float>("scale"), 1.0);
 }
 
 TEST(OpRegistry, CustomChecker) {
@@ -172,6 +172,6 @@ TEST(OpRegistry, CustomChecker) {
   paddle::platform::CPUDeviceContext dev_ctx;
   paddle::framework::Scope scope;
   op->Run(scope, dev_ctx);
-  int test_attr = op->GetAttr<int>("test_attr");
+  int test_attr = op->Attr<int>("test_attr");
   ASSERT_EQ(test_attr, 4);
 }
\ No newline at end of file

paddle/framework/operator.cc

@@ -123,6 +123,15 @@ OperatorBase::OperatorBase(const std::string& type,
   CheckAllInputOutputSet();
 }
 
+std::vector<std::string> OperatorBase::InputVars() const {
+  std::vector<std::string> ret_val;
+  for (auto& o : outputs_) {
+    ret_val.reserve(ret_val.size() + o.second.size());
+    ret_val.insert(ret_val.end(), o.second.begin(), o.second.end());
+  }
+  return ret_val;
+}
+
 std::vector<std::string> OperatorBase::OutputVars(bool has_intermediate) const {
   std::vector<std::string> ret_val;
   if (has_intermediate) {

paddle/framework/operator.h

@@ -69,7 +69,7 @@ class OperatorBase {
   virtual ~OperatorBase() {}
 
   template <typename T>
-  inline const T& GetAttr(const std::string& name) const {
+  inline const T& Attr(const std::string& name) const {
     PADDLE_ENFORCE(attrs_.count(name) != 0, "%s should be in AttributeMap",
                    name);
     return boost::get<T>(attrs_.at(name));
@@ -94,11 +94,14 @@ class OperatorBase {
 
   const VariableNameMap& Inputs() const { return inputs_; }
   const VariableNameMap& Outputs() const { return outputs_; }
+
   //! Get a input with argument's name described in `op_proto`
   std::string Input(const std::string& name) const;
   //! Get a input which has multiple variables.
   const std::vector<std::string>& Inputs(const std::string& name) const;
 
+  std::vector<std::string> InputVars() const;
+
   //! Get a output with argument's name described in `op_proto`
   std::string Output(const std::string& name) const;
   //! Get an output which has multiple variables.
@@ -238,8 +241,8 @@ class InferShapeContext {
   const Scope& scope() const { return scope_; }
 
   template <typename T>
-  inline const T& GetAttr(const std::string& name) const {
-    return op_.GetAttr<T>(name);
+  inline const T& Attr(const std::string& name) const {
+    return op_.Attr<T>(name);
   }
 
   size_t InputSize(const std::string& name) const {
@@ -311,9 +314,9 @@ class InferShapeContext {
   }
 
   template <typename T>
-  std::vector<const T*> MultiOutput(const std::string& name) const {
+  std::vector<T*> MultiOutput(const std::string& name) const {
     auto names = op_.Outputs(name);
-    std::vector<const T*> res;
+    std::vector<T*> res;
     res.reserve(names.size());
     std::transform(names.begin(), names.end(), std::back_inserter(res),
                    [&](const std::string& sub_name) {

paddle/framework/operator_test.cc

@@ -102,7 +102,7 @@ class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
     AddOutput("y", "output of test op");
     AddAttr<float>("scale", "scale of cosine op")
         .SetDefault(1.0)
-        .LargerThan(0.0);
+        .GreaterThan(0.0);
     AddComment("This is test op");
   }
 };
@@ -140,7 +140,7 @@ class OpKernelTestMultiInputsProtoAndCheckerMaker
     AddOutput("ys", "outputs of test op").AsDuplicable();
     AddAttr<float>("scale", "scale of cosine op")
         .SetDefault(1.0)
-        .LargerThan(0.0);
+        .GreaterThan(0.0);
     AddComment("This is test op");
   }
 };

paddle/framework/tensor.h

@@ -43,6 +43,9 @@ class Tensor {
   template <typename T, size_t D, int MajorType, typename IndexType>
   friend struct EigenTensor;
 
+  template <typename T, int MajorType, typename IndexType>
+  friend struct EigenMatrix;
+
   template <typename T, int MajorType, typename IndexType>
   friend struct EigenVector;
 

paddle/framework/tensor_impl.h

@@ -148,5 +148,13 @@ inline Tensor& Tensor::Resize(const DDim& dims) {
 
 inline const DDim& Tensor::dims() const { return dims_; }
 
+template <typename T>
+inline Tensor ReshapeToMatrix(const Tensor& src, int num_col_dims) {
+  Tensor res;
+  res.ShareDataWith<T>(src);
+  res.Resize(flatten_to_2d(src.dims(), num_col_dims));
+  return res;
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/tensor_test.cc

@@ -262,3 +262,16 @@ TEST(Tensor, CopyFrom) {
   }
 #endif
 }
+
+TEST(Tensor, ReshapeToMatrix) {
+  using namespace paddle::framework;
+  using namespace paddle::platform;
+  Tensor src;
+  int* src_ptr = src.mutable_data<int>({2, 3, 4, 9}, CPUPlace());
+  for (int i = 0; i < 2 * 3 * 4 * 9; ++i) {
+    src_ptr[i] = i;
+  }
+  Tensor res = ReshapeToMatrix<int>(src, 2);
+  ASSERT_EQ(res.dims()[0], 2 * 3);
+  ASSERT_EQ(res.dims()[1], 4 * 9);
+}
\ No newline at end of file

paddle/function/CMakeLists.txt

@@ -44,6 +44,7 @@ if(WITH_GPU)
     add_simple_unittest(RowConvOpTest)
     add_simple_unittest(BlockExpandOpTest)
     add_simple_unittest(CropOpTest)
+    add_simple_unittest(SwitchOpTest)
 endif()
 
 add_simple_unittest(Im2ColTest)

paddle/function/EigenGemm.cpp

@@ -83,9 +83,9 @@ struct EigenBlasGemm {
 };
 
 #ifdef PADDLE_TYPE_DOUBLE
-template class EigenBlasGemm<double>;
+template struct EigenBlasGemm<double>;
 #else
-template class EigenBlasGemm<float>;
+template struct EigenBlasGemm<float>;
 #endif
 
 }  // namespace paddle

paddle/function/GruFunctor.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 "GemmFunctor.h"
+#include "hl_cpu_gru.cuh"
+
+namespace paddle {
+
+template <DeviceType Device, class T>
+struct GruFunctor {
+  template <class OpResetOutput, class OpFinalOutput>
+  static void compute(OpResetOutput opResetOutput,
+                      OpFinalOutput opFinalOutput,
+                      hl_gru_value value,
+                      int frameSize,
+                      int batchSize,
+                      hl_activation_mode_t active_node,
+                      hl_activation_mode_t active_gate) {
+#ifndef __NVCC__
+    if (value.prevOutValue) {
+      BlasGemm<Device, T>::compute(false,
+                                   false,
+                                   batchSize,
+                                   2 * frameSize,
+                                   frameSize,
+                                   1,
+                                   value.prevOutValue,
+                                   frameSize,
+                                   value.gateWeight,
+                                   frameSize * 2,
+                                   1,
+                                   value.gateValue,
+                                   frameSize * 3);
+    }
+
+    forward_reset_output(
+        opResetOutput, value, frameSize, batchSize, active_gate);
+
+    if (value.prevOutValue) {
+      BlasGemm<Device, T>::compute(false,
+                                   false,
+                                   batchSize,
+                                   frameSize,
+                                   frameSize,
+                                   1,
+                                   value.resetOutputValue,
+                                   frameSize,
+                                   value.stateWeight,
+                                   frameSize,
+                                   1,
+                                   value.gateValue + frameSize * 2,
+                                   frameSize * 3);
+    }
+
+    forward_final_output(
+        opFinalOutput, value, frameSize, batchSize, active_node);
+#endif
+  }
+};
+
+template <DeviceType Device, class T>
+struct GruGradFunctor {
+  template <class OpStateGrad, class OpResetGrad>
+  static void compute(OpStateGrad opStateGrad,
+                      OpResetGrad opResetGrad,
+                      hl_gru_value value,
+                      hl_gru_grad grad,
+                      int frameSize,
+                      int batchSize,
+                      hl_activation_mode_t active_node,
+                      hl_activation_mode_t active_gate) {
+#ifndef __NVCC__
+    backward_state_grad(
+        opStateGrad, value, grad, frameSize, batchSize, active_node);
+
+    if (value.prevOutValue && grad.prevOutGrad) {
+      BlasGemm<Device, T>::compute(false,
+                                   true,
+                                   batchSize,
+                                   frameSize,
+                                   frameSize,
+                                   1,
+                                   grad.gateGrad + frameSize * 2,
+                                   frameSize * 3,
+                                   value.stateWeight,
+                                   frameSize,
+                                   0,
+                                   grad.resetOutputGrad,
+                                   frameSize);
+
+      if (grad.stateWeightGrad) {
+        BlasGemm<Device, T>::compute(true,
+                                     false,
+                                     frameSize,
+                                     frameSize,
+                                     batchSize,
+                                     1,
+                                     value.resetOutputValue,
+                                     frameSize,
+                                     grad.gateGrad + frameSize * 2,
+                                     frameSize * 3,
+                                     1,
+                                     grad.stateWeightGrad,
+                                     frameSize);
+      }
+    }
+
+    backward_reset_grad(
+        opResetGrad, value, grad, frameSize, batchSize, active_gate);
+
+    if (grad.prevOutGrad && value.prevOutValue) {
+      BlasGemm<Device, T>::compute(false,
+                                   true,
+                                   batchSize,
+                                   frameSize,
+                                   frameSize * 2,
+                                   1,
+                                   grad.gateGrad,
+                                   frameSize * 3,
+                                   value.gateWeight,
+                                   frameSize * 2,
+                                   1,
+                                   grad.prevOutGrad,
+                                   frameSize);
+
+      if (grad.gateWeightGrad) {
+        BlasGemm<Device, T>::compute(true,
+                                     false,
+                                     frameSize,
+                                     frameSize * 2,
+                                     batchSize,
+                                     1,
+                                     value.prevOutValue,
+                                     frameSize,
+                                     grad.gateGrad,
+                                     frameSize * 3,
+                                     1,
+                                     grad.gateWeightGrad,
+                                     frameSize * 2);
+      }
+    }
+#endif
+  }
+};
+
+}  // namespace paddle

paddle/function/Im2Col.h

@@ -94,95 +94,4 @@ 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/MulOp.cpp

@@ -13,18 +13,10 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "MulOp.h"
-/// todo(tianbing), delete it
-#include <iostream>
-#include "paddle/math/MathFunctions.h"
+#include "GemmFunctor.h"
 #include "paddle/math/SIMDFunctions.h"
 #include "paddle/utils/ThreadLocal.h"
 
-#ifndef PADDLE_TYPE_DOUBLE
-#define GEMM paddle::gemm<float>
-#else
-#define GEMM paddle::gemm<double>
-#endif
-
 namespace {
 inline void vecAddTo(real* a, const real* b, real scaleB, size_t len) {
   for (unsigned int i = 0; i < len; ++i) {
@@ -114,19 +106,20 @@ void MulOp<DEVICE_TYPE_CPU>(CpuMatrix& out,
                             real scaleT,
                             bool aTrans,
                             bool bTrans) {
-  GEMM(aTrans ? CblasTrans : CblasNoTrans,
-       bTrans ? CblasTrans : CblasNoTrans,
-       out.getHeight(),
-       out.getWidth(),
-       !aTrans ? a.getWidth() : a.getHeight(),
-       scaleAB,
-       a.getData(),
-       a.getStride(),
-       b.getData(),
-       b.getStride(),
-       scaleT,
-       out.getData(),
-       out.getStride());
+  BlasGemm<DEVICE_TYPE_CPU, real>::compute(
+      aTrans,
+      bTrans,
+      out.getHeight(),
+      out.getWidth(),
+      !aTrans ? a.getWidth() : a.getHeight(),
+      scaleAB,
+      a.getData(),
+      a.getStride(),
+      b.getData(),
+      b.getStride(),
+      scaleT,
+      out.getData(),
+      out.getStride());
 }
 
 /// dense matrix (+)= sparse matrix * dense matrix

paddle/function/SwitchOp.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 "SwitchOp.h"
+#include "paddle/math/Vector.h"
+
+namespace paddle {
+
+template <>
+void NCHW2NHWC<DEVICE_TYPE_CPU>(real* outputs,
+                                const real* inputs,
+                                const int num,
+                                const int inC,
+                                const int inH,
+                                const int inW,
+                                const int argType) {
+  for (int n = 0; n < num; ++n) {
+    for (int c = 0; c < inC; ++c) {
+      for (int h = 0; h < inH; ++h) {
+        for (int w = 0; w < inW; ++w) {
+          if (argType == ADD_TO) {
+            outputs[((n * inH + h) * inW + w) * inC + c] += *(inputs++);
+          } else {
+            outputs[((n * inH + h) * inW + w) * inC + c] = *(inputs++);
+          }
+        }
+      }
+    }
+  }
+}
+
+template <>
+void NHWC2NCHW<DEVICE_TYPE_CPU>(real* outputs,
+                                const real* inputs,
+                                const int num,
+                                const int inH,
+                                const int inW,
+                                const int inC,
+                                const int argType) {
+  for (int n = 0; n < num; ++n) {
+    for (int h = 0; h < inH; ++h) {
+      for (int w = 0; w < inW; ++w) {
+        for (int c = 0; c < inC; ++c) {
+          if (argType == ADD_TO) {
+            outputs[((n * inC + c) * inH + h) * inW + w] += *(inputs++);
+          } else {
+            outputs[((n * inC + c) * inH + h) * inW + w] = *(inputs++);
+          }
+        }
+      }
+    }
+  }
+}
+
+/**
+ * \brief  Switch dimension order of image input.
+ *         The input and output is a 4D tensor. Switch order
+ *         'batch_size,channels, height, width' to
+ *         order 'batch_size, height, width, channels'.
+ *
+ * Argument in this Function:
+ * \param inputs  input data with order 'batch_size,channels, height, width'.
+ * \param outputs output data with order 'batch_size, height, width, channels'.
+ */
+template <DeviceType Device>
+class NCHW2NHWCFunc : public FunctionBase {
+public:
+  void init(const FuncConfig& config) override {}
+
+  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
+    CHECK_EQ(1UL, inputs.size());
+    CHECK_EQ(1UL, outputs.size());
+
+    size_t num = inputs[0].shape()[0];
+    size_t inC = inputs[0].shape()[1];
+    size_t inH = inputs[0].shape()[2];
+    size_t inW = inputs[0].shape()[3];
+    NCHW2NHWC<Device>(outputs[0].data<real>(),
+                      inputs[0].data<real>(),
+                      num,
+                      inC,
+                      inH,
+                      inW,
+                      outputs[0].getArgType());
+  }
+};
+
+/**
+ * \brief  Switch dimension order of image input.
+ *         The input and output is a 4D tensor. Switch order
+ *         'batch_size, height, width, channels' to
+ *         order 'batch_size, channels, height, width'.
+ *
+ * Argument in this Function:
+ * \param inputs  input data with order 'batch_size, height, width, channels'.
+ * \param outputs output data with order 'batch_size, channels, height, width'.
+ */
+template <DeviceType Device>
+class NHWC2NCHWFunc : public FunctionBase {
+public:
+  void init(const FuncConfig& config) override {}
+
+  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
+    CHECK_EQ(1UL, inputs.size());
+    CHECK_EQ(1UL, outputs.size());
+
+    size_t num = inputs[0].shape()[0];
+    size_t inH = inputs[0].shape()[1];
+    size_t inW = inputs[0].shape()[2];
+    size_t inC = inputs[0].shape()[3];
+
+    NHWC2NCHW<Device>(outputs[0].data<real>(),
+                      inputs[0].data<real>(),
+                      num,
+                      inH,
+                      inW,
+                      inC,
+                      outputs[0].getArgType());
+  }
+};
+
+REGISTER_TYPED_FUNC(NCHW2NHWC, CPU, NCHW2NHWCFunc);
+REGISTER_TYPED_FUNC(NHWC2NCHW, CPU, NHWC2NCHWFunc);
+#ifndef PADDLE_ONLY_CPU
+REGISTER_TYPED_FUNC(NCHW2NHWC, GPU, NCHW2NHWCFunc);
+REGISTER_TYPED_FUNC(NHWC2NCHW, GPU, NHWC2NCHWFunc);
+#endif
+
+}  // namespace paddle

paddle/function/SwitchOp.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include "Function.h"
+
+namespace paddle {
+
+/**
+ * \brief  This funtion switch dimension order of image input.
+ *         The input and output is a 4D tensor. Switch order 'batch_size,
+ *channels, height, width' to
+ *         order 'batch_size, height, width, channels'.
+ *
+ * \param[out] outputs save results.
+ * \param[in]  inputs  input data.
+ * \param[in]  num     batch size of input data.
+ * \param[in]  inC     channel number of input data.
+ * \param[in]  inH     height of input data.
+ * \param[in]  inH     with of input data.
+ * \param[in]  argType     type of output argument.
+ */
+template <DeviceType Device>
+void NCHW2NHWC(real* outputs,
+               const real* inputs,
+               const int num,
+               const int inC,
+               const int inH,
+               const int inW,
+               const int argtype);
+
+/**
+ * \brief  This funtion switch dimension order of image input.
+ *         The input and output is a 4D tensor. Switch order 'batch_size,
+ *height, width, channels' to
+ *         order 'batch_size, channels, height, width'.
+ *
+ * \param[out] inGrad  gradients of previous layer.
+ * \param[in]  outGrad output gradients.
+ * \param[in]  num     batch size of input data.
+ * \param[in]  inH     height of input data.
+ * \param[in]  inW     with of input data.
+ * \param[in]  inC     channel number of input data.
+ * \param[in]  argType     type of output argument.
+ */
+template <DeviceType Device>
+void NHWC2NCHW(real* inGrad,
+               const real* outGrad,
+               const int num,
+               const int inH,
+               const int inW,
+               const int inC,
+               const int argType);
+}  // namespace paddle

paddle/function/SwitchOpGpu.cu

+/* Copyright (c) 2016 Paddle
+
+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 "SwitchOp.h"
+#include "hl_base.h"
+
+namespace paddle {
+
+__global__ void KeNCHW2NHWC(real* outputs,
+                            const real* inputs,
+                            int inC,
+                            int inH,
+                            int inW,
+                            int nthreads,
+                            int argType) {
+  const int idx = threadIdx.x + blockIdx.x * blockDim.x;
+  if (idx < nthreads) {
+    const int w = idx % inW;
+    const int h = (idx / inW) % inH;
+    const int c = (idx / inW / inH) % inC;
+    const int n = idx / inW / inH / inC;
+
+    const int off = ((n * inH + h) * inW + w) * inC + c;
+    if (argType == ADD_TO) {
+      outputs[off] += inputs[idx];
+    } else {
+      outputs[off] = inputs[idx];
+    }
+  }
+}
+
+template <>
+void NCHW2NHWC<DEVICE_TYPE_GPU>(real* outputs,
+                                const real* inputs,
+                                const int num,
+                                const int inC,
+                                const int inH,
+                                const int inW,
+                                const int argType) {
+  size_t nth = num * inC * inH * inW;
+  int blockSize = 1024;
+  int gridSize = (nth + 1024 - 1) / 1024;
+  KeNCHW2NHWC<<<gridSize, blockSize, 0, STREAM_DEFAULT>>>(
+      outputs, inputs, inC, inH, inW, nth, argType);
+  CHECK_SYNC("NCHW2NHWC");
+}
+
+__global__ void KeNHWC2NCHW(real* outputs,
+                            const real* inputs,
+                            int inH,
+                            int inW,
+                            int inC,
+                            int nthreads,
+                            int argType) {
+  const int idx = threadIdx.x + blockIdx.x * blockDim.x;
+  if (idx < nthreads) {
+    const int c = idx % inC;
+    const int w = (idx / inC) % inW;
+    const int h = (idx / inC / inW) % inH;
+    const int n = idx / inW / inH / inC;
+
+    const int off = ((n * inC + c) * inH + h) * inW + w;
+    if (argType == ADD_TO) {
+      outputs[off] += inputs[idx];
+    } else {
+      outputs[off] = inputs[idx];
+    }
+  }
+}
+
+template <>
+void NHWC2NCHW<DEVICE_TYPE_GPU>(real* outputs,
+                                const real* inputs,
+                                const int num,
+                                const int inH,
+                                const int inW,
+                                const int inC,
+                                const int argType) {
+  int nth = num * inC * inH * inW;
+  int blockSize = 1024;
+  int gridSize = (nth + 1024 - 1) / 1024;
+  KeNHWC2NCHW<<<gridSize, blockSize, 0, STREAM_DEFAULT>>>(
+      outputs, inputs, inH, inW, inC, nth, argType);
+  CHECK_SYNC("NHWC2NCHW");
+}
+
+}  // namespace paddle

paddle/function/SwitchOpTest.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <gtest/gtest.h>
+#include "FunctionTest.h"
+
+namespace paddle {
+
+TEST(Pad, real) {
+  for (size_t numSamples : {1, 4, 8, 16}) {
+    for (size_t channels : {1, 4, 8, 16}) {
+      for (size_t imgSizeH : {1, 4, 8, 16}) {
+        for (size_t imgSizeW : {1, 4, 8, 16}) {
+          VLOG(3) << " numSamples=" << numSamples << " channels=" << channels
+                  << " imgSizeH=" << imgSizeH << " imgSizeW=" << imgSizeW;
+          for (bool test_grad : {true, false}) {
+            CpuGpuFuncCompare compare(test_grad ? "NHWC2NCHW" : "NCHW2NHWC",
+                                      FuncConfig());
+            TensorShape inDims{numSamples, channels, imgSizeH, imgSizeW};
+            TensorShape outDims{numSamples, imgSizeH, imgSizeW, channels};
+            compare.addInputs(
+                BufferArg(VALUE_TYPE_FLOAT, test_grad ? outDims : inDims));
+            compare.addOutputs(BufferArg(
+                VALUE_TYPE_FLOAT, test_grad ? inDims : outDims, ASSIGN_TO));
+            compare.run();
+          }
+        }
+      }
+    }
+  }
+}
+
+}  // namespace paddle

paddle/function/neon/NeonDepthwiseConvTranspose.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 "NeonDepthwiseConv.h"
+#include "paddle/function/ConvOp.h"
+
+namespace paddle {
+
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+
+template <DeviceType Device>
+class NeonDepthwiseConvTransposeFunction : 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();
+
+    int batchSize = input[0];
+    int inputChannels = input[1];
+    int inputHeight = input[2];
+    int inputWidth = input[3];
+    int filterHeight = getFilterHeight(filter);
+    int filterWidth = getFilterWidth(filter);
+    int outputChannels = output[1];
+    int outputHeight = output[2];
+    int outputWidth = output[3];
+    int filterMultiplier = outputChannels / groups_;
+    CHECK_EQ(inputChannels, groups_);
+
+    // only support strideH() == strideW() and filterHeight == filterWidth.
+    CHECK_EQ(strideH(), strideW());
+    CHECK_EQ(paddingH(), paddingW());
+    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, input -> inputPadding
+    float* inputPadding = inputData;
+    int padInputHeight =
+        (inputHeight - 1) * strideH() + 2 * filterHeight - 1 - 2 * paddingH();
+    int padInputWidth =
+        (inputWidth - 1) * strideW() + 2 * filterWidth - 1 - 2 * paddingW();
+
+    if (padInputHeight > inputHeight || padInputWidth > inputWidth) {
+      int newSize = batchSize * inputChannels * padInputHeight * padInputWidth;
+      resizeBuffer<Device>(newSize);
+      inputPadding = reinterpret_cast<float*>(memory_->getBuf());
+      if (strideH() == 1) {
+        neon::Padding<float>::run(inputData,
+                                  inputPadding,
+                                  batchSize * inputChannels,
+                                  inputHeight,
+                                  inputWidth,
+                                  padInputHeight,
+                                  padInputWidth);
+      } else if (strideH() == 2) {
+        neon::StridePadding::run(inputData,
+                                 inputPadding,
+                                 batchSize * inputChannels,
+                                 inputHeight,
+                                 inputWidth,
+                                 padInputHeight,
+                                 padInputWidth);
+      } else {
+        LOG(FATAL) << "Not supported";
+      }
+    }
+
+    std::function<void(
+        const float*, const float*, int, int, int, int, int, int, float*)>
+        DepthWiseConv;
+
+    if (filterWidth == 3) {
+      DepthWiseConv = neon::DepthwiseConvKernel<3, 1>::run;
+    } else if (filterWidth == 4) {
+      DepthWiseConv = neon::DepthwiseConvKernel<4, 1>::run;
+    } else {
+      LOG(FATAL) << "Not supported";
+    }
+
+    for (int i = 0; i < batchSize; i++) {
+      DepthWiseConv(inputPadding,
+                    filterData,
+                    padInputHeight,
+                    padInputWidth,
+                    outputChannels,
+                    outputHeight,
+                    outputWidth,
+                    filterMultiplier,
+                    outputData);
+      inputPadding += inputChannels * padInputHeight * padInputWidth;
+      outputData += outputChannels * outputHeight * outputWidth;
+    }
+  }
+};
+
+#ifndef PADDLE_TYPE_DOUBLE
+
+REGISTER_TYPED_FUNC(NeonDepthwiseConvTranspose,
+                    CPU,
+                    NeonDepthwiseConvTransposeFunction);
+
+#endif
+
+#endif
+
+}  // namespace paddle

paddle/function/neon/neon_util.h

@@ -33,12 +33,8 @@ inline float32_t vaddvq_f32(float32x4_t 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));
-}
+#define vmlaq_laneq_f32(a, b, v, lane) \
+  vmlaq_n_f32(a, b, vgetq_lane_f32(v, lane))
 #endif
 
 }  // namespace neon

paddle/gserver/layers/BatchNormBaseLayer.cpp

@@ -62,14 +62,18 @@ void BatchNormBaseLayer::calFeatureMapSize() {
   const ImageConfig& conf = config_.inputs(0).image_conf();
   imageH_ = inputLayers_[0]->getOutput().getFrameHeight();
   imageW_ = inputLayers_[0]->getOutput().getFrameWidth();
+  imageD_ = inputLayers_[0]->getOutput().getFrameDepth();
+
+  if (0 == imageD_) imageD_ = conf.img_size_z();
   if (imageH_ == 0 && imageW_ == 0) {
     imageH_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size();
     imageW_ = conf.img_size();
   } else {
     getOutput().setFrameHeight(imageH_);
     getOutput().setFrameWidth(imageW_);
+    getOutput().setFrameDepth(imageD_);
   }
-  imgPixels_ = imageH_ * imageW_;
+  imgPixels_ = imageH_ * imageW_ * imageD_;
 }
 
 }  // namespace paddle

paddle/gserver/layers/BatchNormBaseLayer.h

@@ -80,6 +80,7 @@ protected:
 
   /// Height or width of input image feature.
   /// Both of them are 1 if the input is fully-connected layer.
+  int imageD_;
   int imageH_;
   int imageW_;
   /// Height * Width.

paddle/gserver/layers/Conv3DLayer.cpp

@@ -83,8 +83,8 @@ void Conv3DLayer::forward(PassType passType) {
   int outWidth = getSize();
   resetOutput(batchSize, outWidth);
 
+  REGISTER_TIMER_INFO("FwdConv3D", getName().c_str());
   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];
@@ -120,7 +120,6 @@ void Conv3DLayer::forward(PassType passType) {
     }
   }
   if (nullptr != this->biasParameter_) {
-    REGISTER_TIMER_INFO("FwBiasTimer", getName().c_str());
     this->addBias();
   }
   forwardActivation();
@@ -134,15 +133,14 @@ void Conv3DLayer::backward(const UpdateCallback &callback) {
     biases_->getParameterPtr()->incUpdate(callback);
   }
 
+  REGISTER_TIMER_INFO("BwdConv3D", getName().c_str());
   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);
   }
 }

paddle/gserver/layers/CudnnBatchNormLayer.cpp

@@ -37,7 +37,7 @@ bool CudnnBatchNormLayer::init(const LayerMap& layerMap,
 }
 
 void CudnnBatchNormLayer::reshape(int batchSize) {
-  hl_tensor_reshape(ioDesc_, batchSize, channels_, imageH_, imageW_);
+  hl_tensor_reshape(ioDesc_, batchSize, channels_, imageH_ * imageD_, imageW_);
 }
 
 void CudnnBatchNormLayer::forward(PassType passType) {
@@ -104,7 +104,7 @@ void CudnnBatchNormLayer::forward(PassType passType) {
                                    EPS,
                                    batchSize,
                                    channels_,
-                                   imageH_,
+                                   imageH_ * imageD_,
                                    imageW_);
     }
   }

paddle/gserver/layers/DeConv3DLayer.cpp

@@ -84,8 +84,8 @@ void DeConv3DLayer::forward(PassType passType) {
   resetOutput(batchSize, outWidth);
   const MatrixPtr outMat = getOutputValue();
 
+  REGISTER_TIMER_INFO("FwdDeConv3D", getName().c_str());
   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];
@@ -120,7 +120,6 @@ void DeConv3DLayer::forward(PassType passType) {
     }
   }
   if (nullptr != this->biasParameter_) {
-    REGISTER_TIMER_INFO("FwBiasTimer", getName().c_str());
     this->addBias();
   }
   forwardActivation();
@@ -133,12 +132,12 @@ void DeConv3DLayer::backward(const UpdateCallback &callback) {
     bpropBiases();
     biases_->getParameterPtr()->incUpdate(callback);
   }
+  REGISTER_TIMER_INFO("BwdDeConv3D", getName().c_str());
   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) {
@@ -182,7 +181,6 @@ void DeConv3DLayer::backward(const UpdateCallback &callback) {
           }
         }
       }
-      REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
       weights_[i]->getParameterPtr()->incUpdate(callback);
     }
   }

paddle/gserver/layers/DetectionOutputLayer.cpp

@@ -139,7 +139,13 @@ void DetectionOutputLayer::forward(PassType passType) {
                                        allDecodedBBoxes,
                                        &allIndices);
 
-  resetOutput(numKept, 7);
+  if (numKept > 0) {
+    resetOutput(numKept, 7);
+  } else {
+    MatrixPtr outV = getOutputValue();
+    outV = NULL;
+    return;
+  }
   MatrixPtr outV = getOutputValue();
   getDetectionOutput(confBuffer_->getData(),
                      numKept,

paddle/gserver/layers/DetectionUtil.cpp

@@ -469,7 +469,7 @@ size_t getDetectionIndices(
     const size_t numClasses,
     const size_t backgroundId,
     const size_t batchSize,
-    const size_t confThreshold,
+    const real confThreshold,
     const size_t nmsTopK,
     const real nmsThreshold,
     const size_t keepTopK,

paddle/gserver/layers/DetectionUtil.h

@@ -275,7 +275,7 @@ size_t getDetectionIndices(
     const size_t numClasses,
     const size_t backgroundId,
     const size_t batchSize,
-    const size_t confThreshold,
+    const real confThreshold,
     const size_t nmsTopK,
     const real nmsThreshold,
     const size_t keepTopK,

paddle/gserver/layers/GruCompute.cpp

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #include "GruCompute.h"
 #include "hl_recurrent_apply.cuh"
+#include "paddle/function/GruFunctor.h"
 #include "paddle/utils/Util.h"
 
 namespace paddle {
@@ -25,13 +26,13 @@ void GruCompute::init(LayerConfig &config) {
 
 template <>
 void GruCompute::forward<0>(hl_gru_value value, int frameSize, int batchSize) {
-  hl_cpu_gru_forward(hppl::forward::gru_resetOutput(),
-                     hppl::forward::gru_finalOutput(),
-                     value,
-                     frameSize,
-                     batchSize,
-                     activeNode_,
-                     activeGate_);
+  GruFunctor<DEVICE_TYPE_CPU, real>::compute(hppl::forward::gru_resetOutput(),
+                                             hppl::forward::gru_finalOutput(),
+                                             value,
+                                             frameSize,
+                                             batchSize,
+                                             activeNode_,
+                                             activeGate_);
 }
 
 template <>
@@ -39,14 +40,15 @@ void GruCompute::backward<0>(hl_gru_value value,
                              hl_gru_grad grad,
                              int frameSize,
                              int batchSize) {
-  hl_cpu_gru_backward(hppl::backward::gru_stateGrad(),
-                      hppl::backward::gru_resetGrad(),
-                      value,
-                      grad,
-                      frameSize,
-                      batchSize,
-                      activeNode_,
-                      activeGate_);
+  GruGradFunctor<DEVICE_TYPE_CPU, real>::compute(
+      hppl::backward::gru_stateGrad(),
+      hppl::backward::gru_resetGrad(),
+      value,
+      grad,
+      frameSize,
+      batchSize,
+      activeNode_,
+      activeGate_);
 }
 
 }  // namespace paddle

paddle/gserver/layers/SwitchOrderLayer.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 "SwitchOrderLayer.h"
+#include "paddle/utils/Stat.h"
+
+namespace paddle {
+
+REGISTER_LAYER(switch_order, SwitchOrderLayer);
+
+bool SwitchOrderLayer::init(const LayerMap& layerMap,
+                            const ParameterMap& parameterMap) {
+  /* Initialize the basic parent class */
+  Layer::init(layerMap, parameterMap);
+  auto& img_conf = config_.inputs(0).image_conf();
+  size_t inD = img_conf.img_size_z();
+  size_t inH =
+      img_conf.has_img_size_y() ? img_conf.img_size_y() : img_conf.img_size();
+  size_t inW = img_conf.img_size();
+  size_t inC = img_conf.channels();
+  inH = inH * inD;
+  inDims_ = TensorShape({0, inC, inH, inW});
+  outDims_ = TensorShape(4);
+
+  auto& reshape_conf = config_.reshape_conf();
+  for (int i = 0; i < reshape_conf.height_axis_size(); i++) {
+    heightAxis_.push_back(reshape_conf.height_axis(i));
+  }
+  for (int i = 0; i < reshape_conf.width_axis_size(); i++) {
+    widthAxis_.push_back(reshape_conf.width_axis(i));
+  }
+  createFunction(nchw2nhwc_, "NCHW2NHWC", FuncConfig());
+  createFunction(nhwc2nchw_, "NHWC2NCHW", FuncConfig());
+  return true;
+}
+
+void SwitchOrderLayer::setOutDims() {
+  outDims_.setDim(0, inDims_[0]);
+  outDims_.setDim(1, inDims_[2]);
+  outDims_.setDim(2, inDims_[3]);
+  outDims_.setDim(3, inDims_[1]);
+  reshapeHeight_ = 1;
+  for (size_t i = 0; i < heightAxis_.size(); i++) {
+    reshapeHeight_ *= outDims_[heightAxis_[i]];
+  }
+  output_.setFrameHeight(reshapeHeight_);
+  reshapeWidth_ = 1;
+  for (size_t i = 0; i < widthAxis_.size(); i++) {
+    reshapeWidth_ *= outDims_[widthAxis_[i]];
+  }
+  output_.setFrameWidth(reshapeWidth_);
+}
+
+void SwitchOrderLayer::setInDims() {
+  MatrixPtr input = inputLayers_[0]->getOutputValue();
+  size_t batchSize = input->getHeight();
+  inDims_.setDim(0, batchSize);
+  int d = inputLayers_[0]->getOutput().getFrameDepth();
+  d = (d == 0 ? 1 : d);
+  int h = inputLayers_[0]->getOutput().getFrameHeight();
+  if (h != 0) inDims_.setDim(2, h * d);
+  int w = inputLayers_[0]->getOutput().getFrameWidth();
+  if (w != 0) inDims_.setDim(3, w);
+  int totalCount = input->getElementCnt();
+  int channels = totalCount / (inDims_[0] * inDims_[2] * inDims_[3]);
+  if (channels != 0) inDims_.setDim(1, channels);
+}
+
+void SwitchOrderLayer::forward(PassType passType) {
+  Layer::forward(passType);
+  setInDims();
+  setOutDims();
+  resetOutput(outDims_[0], outDims_[1] * outDims_[2] * outDims_[3]);
+  if (heightAxis_.size() > 0) {
+    getOutputValue()->reshape(reshapeHeight_, reshapeWidth_);
+    getOutputGrad()->reshape(reshapeHeight_, reshapeWidth_);
+  }
+
+  // switch NCHW to NHWC
+  BufferArgs inputs;
+  BufferArgs outputs;
+  inputs.addArg(*getInputValue(0), inDims_);
+  outputs.addArg(*getOutputValue(), outDims_);
+  nchw2nhwc_[0]->calc(inputs, outputs);
+  forwardActivation();
+}
+
+void SwitchOrderLayer::backward(const UpdateCallback& callback) {
+  (void)callback;
+  backwardActivation();
+
+  // switch NHWC to NCHW
+  BufferArgs inputs;
+  BufferArgs outputs;
+  inputs.addArg(*getOutputGrad(), outDims_);
+  outputs.addArg(*getInputGrad(0), inDims_, ADD_TO);
+  nhwc2nchw_[0]->calc(inputs, outputs);
+}
+}  // namespace paddle

paddle/gserver/layers/SwitchOrderLayer.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include "Layer.h"
+
+namespace paddle {
+
+/**
+ * \brief  This layer calculate softmax in image channel dimension.
+ */
+class SwitchOrderLayer : public Layer {
+public:
+  explicit SwitchOrderLayer(const LayerConfig& config) : Layer(config) {}
+
+  ~SwitchOrderLayer() {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
+  void setInDims();
+  void setOutDims();
+
+protected:
+  std::vector<std::shared_ptr<FunctionBase>> nchw2nhwc_;
+  std::vector<std::shared_ptr<FunctionBase>> nhwc2nchw_;
+  TensorShape inDims_;
+  TensorShape outDims_;
+  std::vector<int> heightAxis_;
+  std::vector<int> widthAxis_;
+  size_t reshapeHeight_;
+  size_t reshapeWidth_;
+};
+}  // namespace paddle

paddle/gserver/tests/test_LayerGrad.cpp

@@ -1703,6 +1703,55 @@ TEST(Layer, BatchNormalizationLayer) {
 #endif
 }
 
+void testBatchNorm3DLayer(const string& type, bool trans, bool useGpu) {
+  TestConfig config;
+  const int CHANNELS = 10;
+  const int IMG_SIZE = 16;
+  const int IMG_SIZE_Y = 8;
+  const int IMG_SIZE_Z = 8;
+  size_t size = CHANNELS * IMG_SIZE * IMG_SIZE_Y * IMG_SIZE_Z;
+  config.layerConfig.set_type(type);
+  config.layerConfig.set_size(size);
+  config.layerConfig.set_active_type("sigmoid");
+  config.biasSize = CHANNELS;
+  config.inputDefs.push_back({INPUT_DATA,
+                              "layer_0",
+                              /* dim= */ size,
+                              /* paraSize= */ CHANNELS});
+
+  config.inputDefs.push_back({INPUT_DATA, "layer_1_running_mean", 1, CHANNELS});
+  config.inputDefs.back().isStatic = true;
+  config.inputDefs.push_back({INPUT_DATA, "layer_2_running_var", 1, CHANNELS});
+  config.inputDefs.back().isStatic = true;
+
+  LayerInputConfig* input = config.layerConfig.add_inputs();
+  config.layerConfig.add_inputs();
+  config.layerConfig.add_inputs();
+
+  ImageConfig* img_conf = input->mutable_image_conf();
+  img_conf->set_channels(CHANNELS);
+  img_conf->set_img_size(IMG_SIZE);
+  img_conf->set_img_size_y(IMG_SIZE_Y);
+  img_conf->set_img_size_z(IMG_SIZE_Z);
+
+  testLayerGrad(config,
+                "batch_norm",
+                64,
+                /* trans= */ trans,
+                useGpu,
+                /* useWeight */ true);
+}
+
+TEST(Layer, testBatchNorm3DLayer) {
+  testBatchNorm3DLayer("batch_norm", false, false);
+#ifndef PADDLE_ONLY_CPU
+  testBatchNorm3DLayer("batch_norm", false, true);
+  if (hl_get_cudnn_lib_version() >= int(4000)) {
+    testBatchNorm3DLayer("cudnn_batch_norm", false, true);
+  }
+#endif
+}
+
 void testConvOperator(bool isDeconv) {
   TestConfig config;
   const int NUM_FILTERS = 16;
@@ -2008,6 +2057,31 @@ TEST(Layer, CropLayer) {
   }
 }
 
+TEST(Layer, SwitchOrderLayer) {
+  TestConfig config;
+  // config input_0
+  config.inputDefs.push_back({INPUT_DATA, "layer_0", 1024, 0});
+  LayerInputConfig* input = config.layerConfig.add_inputs();
+  ImageConfig* img = input->mutable_image_conf();
+  img->set_channels(4);
+  img->set_img_size(16);
+  img->set_img_size_y(16);
+
+  ReshapeConfig* reshape = config.layerConfig.mutable_reshape_conf();
+  reshape->add_height_axis(0);
+  reshape->add_height_axis(1);
+  reshape->add_height_axis(2);
+  reshape->add_width_axis(3);
+
+  // config softmax layer
+  config.layerConfig.set_type("switch_order");
+  config.layerConfig.set_name("switchOrderLayer");
+
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config, "switch_order", 100, false, useGpu, true);
+  }
+}
+
 vector<real> randSampling(real range, int n) {
   CHECK_GE(range, n);
   vector<real> num(range);

paddle/math/MathFunctions.cpp

@@ -84,6 +84,7 @@ LAPACK_ROUTINE_EACH(DYNAMIC_LOAD_LAPACK_WRAP)
 
 namespace paddle {
 
+#ifndef PADDLE_USE_EIGEN_FOR_BLAS
 template <>
 void gemm<float>(const CBLAS_TRANSPOSE transA,
                  const CBLAS_TRANSPOSE transB,
@@ -143,6 +144,7 @@ void gemm<double>(const CBLAS_TRANSPOSE transA,
               C,
               ldc);
 }
+#endif
 
 template <>
 int getrf<float>(const CBLAS_ORDER order,
@@ -182,6 +184,7 @@ int getri<double>(const CBLAS_ORDER order,
   return dynload::PADDLE_DGETRI(order, N, A, lda, ipiv);
 }
 
+#ifndef PADDLE_USE_EIGEN_FOR_BLAS
 template <>
 void axpy<float>(const int n, const float alpha, const float* x, float* y) {
   cblas_saxpy(n, alpha, x, 1, y, 1);
@@ -201,6 +204,7 @@ template <>
 double dotProduct<double>(const int n, const double* x, const double* y) {
   return cblas_ddot(n, x, 1, y, 1);
 }
+#endif
 
 #if defined(PADDLE_USE_MKL) || defined(PADDLE_USE_MKLML)
 

paddle/math/MathFunctions.h

@@ -40,7 +40,14 @@ extern "C" {
 
 #ifndef LAPACK_FOUND
 extern "C" {
+#ifndef PADDLE_USE_EIGEN_FOR_BLAS
 #include <cblas.h>
+#else
+typedef enum CBLAS_ORDER {
+  CblasRowMajor = 101,
+  CblasColMajor = 102
+} CBLAS_ORDER;
+#endif
 int LAPACKE_sgetrf(
     int matrix_layout, int m, int n, float* a, int lda, int* ipiv);
 int LAPACKE_dgetrf(
@@ -56,6 +63,7 @@ int LAPACKE_dgetri(
 
 namespace paddle {
 
+#ifndef PADDLE_USE_EIGEN_FOR_BLAS
 template <class T>
 void gemm(const CBLAS_TRANSPOSE transA,
           const CBLAS_TRANSPOSE transB,
@@ -70,6 +78,7 @@ void gemm(const CBLAS_TRANSPOSE transA,
           const T beta,
           T* C,
           const int ldc);
+#endif
 
 template <class T>
 int getrf(const CBLAS_ORDER Order,
@@ -84,10 +93,21 @@ int getri(
     const CBLAS_ORDER Order, const int N, T* A, const int lda, const int* ipiv);
 
 template <class T>
-void axpy(const int n, const T alpha, const T* x, T* y);
+void axpy(const int n, const T alpha, const T* x, T* y) {
+  /// y = y + alpha * x
+  for (int i = 0; i < n; i++) {
+    y[i] = y[i] + alpha * x[i];
+  }
+}
 
 template <class T>
-T dotProduct(const int n, const T* x, const T* y);
+T dotProduct(const int n, const T* x, const T* y) {
+  T result = static_cast<T>(0);
+  for (int i = 0; i < n; i++) {
+    result += x[i] * y[i];
+  }
+  return result;
+}
 
 template <class T>
 void vExp(const int n, const T* a, T* r);

paddle/math/Matrix.cpp

@@ -28,6 +28,7 @@ limitations under the License. */
 #include "hl_top_k.h"
 #include "paddle/utils/Logging.h"
 
+#include "paddle/function/GemmFunctor.h"
 #include "paddle/utils/ThreadLocal.h"
 
 #include "SIMDFunctions.h"
@@ -2773,24 +2774,24 @@ void CpuMatrix::mul(CpuMatrix* a, CpuMatrix* b, real scaleAB, real scaleT) {
   CHECK(!isTransposed()) << "Not supported";
 
   size_t a_col, b_col, a_row, b_row;
-  CBLAS_TRANSPOSE a_trans, b_trans;
+  bool a_trans, b_trans;
   if (!a->isTransposed()) {
     a_col = a->getWidth();
     a_row = a->getHeight();
-    a_trans = CblasNoTrans;
+    a_trans = false;
   } else {
     a_col = a->getHeight();
     a_row = a->getWidth();
-    a_trans = CblasTrans;
+    a_trans = true;
   }
   if (!b->isTransposed()) {
     b_col = b->getWidth();
     b_row = b->getHeight();
-    b_trans = CblasNoTrans;
+    b_trans = false;
   } else {
     b_col = b->getHeight();
     b_row = b->getWidth();
-    b_trans = CblasTrans;
+    b_trans = true;
   }
 
   CHECK_EQ(a_col, b_row);
@@ -2807,7 +2808,7 @@ void CpuMatrix::mul(CpuMatrix* a, CpuMatrix* b, real scaleAB, real scaleT) {
   int lda = a->getStride();
   int ldb = b->getStride();
   int ldc = getStride();
-  gemm<real>(
+  BlasGemm<DEVICE_TYPE_CPU, real>::compute(
       a_trans, b_trans, M, N, K, scaleAB, A, lda, B, ldb, scaleT, C, ldc);
 }
 

paddle/math/Matrix.h

@@ -1616,6 +1616,10 @@ public:
 };
 
 class CpuMatrix : public Matrix {
+private:
+  MatrixPtr sftmaxSum_;
+  MatrixPtr sftmaxDot_;
+
 public:
   CpuMatrix(size_t height, size_t width, bool trans = false);
   CpuMatrix(real* data, size_t height, size_t width, bool trans = false)

paddle/operators/CMakeLists.txt

@@ -14,27 +14,31 @@ function(op_library TARGET)
     cmake_parse_arguments(op_library "${options}" "${oneValueArgs}"
             "${multiValueArgs}" ${ARGN})
 
-    foreach(src ${op_library_SRCS})
-        if (${src} MATCHES ".*\\.cu$")
-            list(APPEND cu_srcs ${src})
-        elseif(${src} MATCHES ".*\\.cc$")
-            list(APPEND cc_srcs ${src})
-        else()
-            message(FATAL_ERROR "${TARGET} Source file ${src} should only be .cc or .cu")
+    list(LENGTH op_library_SRCS op_library_SRCS_len)
+    if (${op_library_SRCS_len} EQUAL 0)
+        if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cc)
+            list(APPEND cc_srcs ${TARGET}.cc)
         endif()
-    endforeach()
+        if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu)
+            list(APPEND cu_srcs ${TARGET}.cu)
+        endif()
+    else()
+        foreach(src ${op_library_SRCS})
+            if (${src} MATCHES ".*\\.cu$")
+                list(APPEND cu_srcs ${src})
+            elseif(${src} MATCHES ".*\\.cc$")
+                list(APPEND cc_srcs ${src})
+            else()
+                message(FATAL_ERROR "${TARGET} Source file ${src} should only be .cc or .cu")
+            endif()
+        endforeach()
+    endif()
 
     list(LENGTH cc_srcs cc_srcs_len)
     if (${cc_srcs_len} EQUAL 0)
         message(FATAL_ERROR "The op library ${TARGET} should contains at least one .cc file")
     endif()
 
-    list(LENGTH cu_srcs cu_srcs_len)
-    list(LENGTH op_library_DEPS dep_len)
-    if (${cu_srcs_len} EQUAL 0 AND ${dep_len} EQUAL 0)
-        message(WARNING "The op library ${TARGET} not support GPU!")
-    endif()
-
     if (WITH_GPU)
         nv_library(${TARGET} SRCS ${cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS}
                 ${op_common_deps})
@@ -46,22 +50,22 @@ endfunction()
 
 add_subdirectory(math)
 
-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)
+set(DEPS_OPS
+    identity_op
+    minus_op
+    mul_op
+    recurrent_op
+    scale_op)
+op_library(identity_op DEPS scale_op)
+op_library(minus_op DEPS scale_op)
+op_library(mul_op DEPS math_function)
 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(scale_op DEPS net_op)
 
+list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS})
 foreach(src ${GENERAL_OPS})
-    op_library(${src} SRCS ${src}.cc ${src}.cu)
+    op_library(${src})
 endforeach()
 
 set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library")

paddle/operators/add_op.cc

@@ -57,7 +57,6 @@ 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, ops::AddOp, ops::AddOpMaker, add_grad, ops::AddOpGrad);
 
-REGISTER_OP_CPU_KERNEL(add_two,
-                       ops::AddKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(add, ops::AddKernel<paddle::platform::CPUPlace, float>);

paddle/operators/add_op.cu

@@ -12,10 +12,7 @@
    See the License for the specific language governing permissions and
    limitations under the License. */
 
-#define EIGEN_USE_GPU
-#include "paddle/framework/op_registry.h"
 #include "paddle/operators/add_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(add_two,
-                       ops::AddKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(add, ops::AddKernel<paddle::platform::GPUPlace, float>);

paddle/operators/cos_sim_op.h

@@ -23,6 +23,9 @@ using Tensor = framework::Tensor;
 template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
 
 template <typename Place, typename T>
 class CosSimKernel : public framework::OpKernel {
@@ -43,14 +46,14 @@ class CosSimKernel : public framework::OpKernel {
     auto new_dims = framework::make_ddim({dims[0], size / dims[0]});
     auto x = EigenMatrix<T>::From(*input_x, new_dims);
     auto y = EigenMatrix<T>::From(*input_y, new_dims);
-    auto z = EigenMatrix<T>::From(*output_z);
-    auto x_norm = EigenMatrix<T>::From(*output_x_norm);
-    auto y_norm = EigenMatrix<T>::From(*output_y_norm);
+    auto z = EigenVector<T>::Flatten(*output_z);
+    auto x_norm = EigenVector<T>::Flatten(*output_x_norm);
+    auto y_norm = EigenVector<T>::Flatten(*output_y_norm);
 
     auto place = context.GetEigenDevice<Place>();
-    auto xy = (x * y).sum(Eigen::array<int, 1>({1}));
-    x_norm.device(place) = x.square().sum(Eigen::array<int, 1>({1})).sqrt();
-    y_norm.device(place) = y.square().sum(Eigen::array<int, 1>({1})).sqrt();
+    auto xy = (x * y).sum(Eigen::array<int, 1>({{1}}));
+    x_norm.device(place) = x.square().sum(Eigen::array<int, 1>({{1}})).sqrt();
+    y_norm.device(place) = y.square().sum(Eigen::array<int, 1>({{1}})).sqrt();
     z.device(place) = xy / x_norm / y_norm;
   }
 };

paddle/operators/crop_op.cc

@@ -26,19 +26,18 @@ class CropOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    auto dim0 = ctx.Input<Tensor>("X")->dims();
+    auto x_dim = ctx.Input<Tensor>("X")->dims();
     auto Y = ctx.Input<Tensor>("Y");
     if (Y == nullptr) {
-      auto shape = GetAttr<std::vector<int>>("shape");
+      auto shape = Attr<std::vector<int>>("shape");
       PADDLE_ENFORCE_EQ(
-          shape.size(), dim0.size(),
+          int64_t(shape.size()), x_dim.size(),
           "Shape size should be equal to dimention size of input tensor.");
       std::vector<int64_t> tensor_shape(shape.size());
-      for (int i = 0; i < shape.size(); ++i) {
+      for (size_t i = 0; i < shape.size(); ++i) {
         tensor_shape[i] = (int64_t)shape[i];
       }
-      ctx.Output<Tensor>("Out")->Resize(
-          paddle::framework::make_ddim(tensor_shape));
+      ctx.Output<Tensor>("Out")->Resize(framework::make_ddim(tensor_shape));
     } else {
       ctx.Output<Tensor>("Out")->Resize(Y->dims());
     }
@@ -49,14 +48,57 @@ class CropOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   CropOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "The input of crop op");
-    AddInput("Y", "The input used as reference for cropping. ");
-    AddOutput("Out", "The output of crop op.");
+    AddInput("X",
+             "The input of pad op. "
+             "The input should be a k-D tensor(k > 0 and k < 7)");
+    AddInput("Y",
+             "The input used as reference for cropping"
+             " with the same dimension as X. ");
+    AddOutput("Out",
+              "The output of crop op "
+              "with the same dimension as X.");
     AddComment(R"DOC(
 Crop Operator.
+Crop input into output, as specified by offsets and shape.
+
+There are two ways to set shape: 
+1. referenc input: crop input X as shape as reference input.
+                    The dimension of reference input should 
+                    be as same as input X.
+2. shape list: crop input X by shape described by a list<int>.
+               The size of shape list should be as same as 
+               dimension size of  input X.
+
+The input should be a k-D tensor(k > 0 and k < 7). As an example:
+
+Given:
+
+X = [[0, 1, 2, 0, 0]
+       [0, 3, 4, 0, 0]
+       [0, 0, 0, 0, 0]]
+
+and 
+
+offsets = [0, 1]
+
+and
+ 
+shape = [2, 2]
+
+then we get 
+
+Out = [[1, 2],
+   [3, 4]]
+
 )DOC");
-    AddAttr<std::vector<int>>("offsets", "The offsets for cropping.");
-    AddAttr<std::vector<int>>("shape", "The shape for cropping.");
+    AddAttr<std::vector<int>>("offsets",
+                              "A list<int> describing offsets to be cropped."
+                              "The size of offsets list should be as same as "
+                              "dimension size of  input X.");
+    AddAttr<std::vector<int>>("shape",
+                              "A list<int> describing the shape of output."
+                              "The size of shape list should be as same as "
+                              "dimension size of  input X.");
   }
 };
 
@@ -76,12 +118,42 @@ class CropOpGrad : public framework::OperatorWithKernel {
   }
 };
 
+template <typename T>
+class CropCPUKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext &context) const override {
+    auto *x = context.Input<Tensor>("X");
+    auto *out = context.Output<Tensor>("Out");
+    auto x_data = x->data<T>();
+    T *out_data = out->mutable_data<T>(paddle::platform::CPUPlace());
+    auto x_dims = x->dims();
+    auto out_dims = out->dims();
+    int64_t out_count = framework::product(out_dims);
+    std::vector<int64_t> x_shape = framework::vectorize(x_dims);
+    std::vector<int64_t> out_shape = framework::vectorize(out_dims);
+
+    auto offsets = context.op().Attr<std::vector<int>>("offsets");
+    PADDLE_ENFORCE_EQ(
+        x_dims.size(), offsets.size(),
+        "Offsets size should be equal to dimension size of input tensor.");
+
+    std::vector<std::pair<int, int>> crop_rules(x_dims.size());
+    for (size_t i = 0; i < crop_rules.size(); ++i) {
+      crop_rules[i].first = offsets[i];
+      crop_rules[i].second = x_dims[i] - out_dims[i] - offsets[i];
+    }
+
+    for (int64_t i = 0; i < out_count; ++i) {
+      out_data[i] = x_data[transIndex(out_shape, x_shape, crop_rules, i)];
+    }
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
 REGISTER_OP(crop, ops::CropOp, ops::CropOpMaker, crop_grad, ops::CropOpGrad);
-REGISTER_OP_CPU_KERNEL(crop,
-                       ops::CropKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(crop, ops::CropCPUKernel<float>);
 REGISTER_OP_CPU_KERNEL(crop_grad,
                        ops::CropGradKernel<paddle::platform::CPUPlace, float>);

paddle/operators/crop_op.cu

@@ -15,8 +15,104 @@
 #define EIGEN_USE_GPU
 #include "paddle/operators/crop_op.h"
 
+#define CUDA_1D_KERNEL_LOOP(i, n)                            \
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < n; \
+       i += blockDim.x * gridDim.x)
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename T, int D>
+__global__ void CropKernel(const int N, const int64_t* out_shape,
+                           const int64_t* x_shape, const int* crop_rules,
+                           const T* x_data, T* out_data) {
+  CUDA_1D_KERNEL_LOOP(index, N) {
+    // int64_t dim_size = out_shape.size();
+    int64_t pos[D];
+
+    for (int64_t i = D - 1; i >= 0; --i) {
+      pos[i] = (index % out_shape[i]) + crop_rules[i * 2];
+      index = index / out_shape[i];
+    }
+
+    int64_t result = pos[0];
+    for (size_t i = 1; i < D; ++i) {
+      result = result * x_shape[i] + pos[i];
+    }
+
+    out_data[index] = x_data[result];
+  }
+}
+
+template <typename T, int D>
+void CropCUDAFunctoin(const framework::ExecutionContext& context) {
+  auto* x = context.Input<Tensor>("X");
+  auto* out = context.Output<Tensor>("Out");
+  auto x_data = x->data<T>();
+  T* out_data = out->mutable_data<T>(paddle::platform::CPUPlace());
+  auto x_dims = x->dims();
+  auto out_dims = out->dims();
+  int64_t out_count = framework::product(out_dims);
+  int64_t* x_shape = &(framework::vectorize(x_dims))[0];
+  int64_t* out_shape = &(framework::vectorize(out_dims))[0];
+
+  auto offsets = context.op().Attr<std::vector<int>>("offsets");
+  PADDLE_ENFORCE_EQ(
+      x_dims.size(), offsets.size(),
+      "Offsets size should be equal to dimension size of input tensor.");
+
+  int crop_rules[D * 2];
+  for (size_t i = 0; i < x_dims.size(); ++i) {
+    crop_rules[i * 2] = offsets[i];
+    crop_rules[i * 2 + 1] = x_dims[i] - out_dims[i] - offsets[i];
+  }
+
+  int n = out_dims[0];
+  int d = out_dims[1];
+  int block = 512;
+  int grid = (n * d + block - 1) / block;
+
+  CropKernel<T, D><<<grid, block>>>(out_count, out_shape, x_shape, crop_rules,
+                                    x_data, out_data);
+}
+
+template <typename T>
+class CropOpCUDAKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    size_t rank = context.Input<Tensor>("X")->dims().size();
+    switch (rank) {
+      case 1:
+        CropCUDAFunctoin<T, 1>(context);
+        break;
+      case 2:
+        CropCUDAFunctoin<T, 2>(context);
+        break;
+      case 3:
+        CropCUDAFunctoin<T, 3>(context);
+        break;
+      case 4:
+        CropCUDAFunctoin<T, 4>(context);
+        break;
+      case 5:
+        CropCUDAFunctoin<T, 5>(context);
+        break;
+      case 6:
+        CropCUDAFunctoin<T, 6>(context);
+        break;
+      default:
+        PADDLE_THROW(
+            "CropOp only support tensors with no more than 6 dimensions.");
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(crop,
-                       ops::CropKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(crop, ops::CropOpCUDAKernel<float>);
 REGISTER_OP_GPU_KERNEL(crop_grad,
                        ops::CropGradKernel<paddle::platform::GPUPlace, float>);

paddle/operators/crop_op.h

@@ -18,7 +18,7 @@
 #include "paddle/framework/op_registry.h"
 
 namespace paddle {
-namespace operators {
+namespace operators {  // Internal
 
 template <typename T, size_t D, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
@@ -26,60 +26,22 @@ using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>;
 
 using Tensor = framework::Tensor;
 
-template <typename Place, typename T, size_t D>
-void CropFunction(const framework::ExecutionContext& context) {
-  auto* x = context.Input<Tensor>("X");
-  auto* out = context.Output<Tensor>("Out");
-  out->mutable_data<T>(context.GetPlace());
-  auto x_dims = x->dims();
-  auto out_dims = out->dims();
-
-  auto offsets = context.op().GetAttr<std::vector<int>>("offsets");
-  PADDLE_ENFORCE_EQ(
-      x_dims.size(), offsets.size(),
-      "Offsets size should be equal to dimension size of input tensor.");
+int64_t transIndex(std::vector<int64_t> out_shape, std::vector<int64_t> x_shape,
+                   std::vector<std::pair<int, int>> crop_rules, size_t index) {
+  int64_t dim_size = out_shape.size();
+  int64_t pos[dim_size];
 
-  Eigen::array<std::pair<int, int>, D> paddings;
-  for (size_t i = 0; i < D; ++i) {
-    paddings[i].first = -(offsets[i]);
-    paddings[i].second = -(x_dims[i] - out_dims[i] - offsets[i]);
+  for (int64_t i = out_shape.size() - 1; i >= 0; --i) {
+    pos[i] = (index % out_shape[i]) + crop_rules[i].first;
+    index = index / out_shape[i];
   }
 
-  auto x_tensor = EigenTensor<T, D>::From(*x);
-  auto out_tensor = EigenTensor<T, D>::From(*out);
-  auto place = context.GetEigenDevice<Place>();
-  out_tensor.device(place) = x_tensor.pad(paddings, 0);
-}
-
-template <typename Place, typename T>
-class CropKernel : public framework::OpKernel {
- public:
-  void Compute(const framework::ExecutionContext& context) const override {
-    int dim = context.Input<Tensor>("X")->dims().size();
-    switch (dim) {
-      case 1:
-        CropFunction<Place, T, 1>(context);
-        break;
-      case 2:
-        CropFunction<Place, T, 2>(context);
-        break;
-      case 3:
-        CropFunction<Place, T, 3>(context);
-        break;
-      case 4:
-        CropFunction<Place, T, 4>(context);
-        break;
-      case 5:
-        CropFunction<Place, T, 5>(context);
-        break;
-      case 6:
-        CropFunction<Place, T, 6>(context);
-        break;
-      default:
-        LOG(ERROR) << "Only ranks up to 6 supported.";
-    }
+  size_t result = pos[0];
+  for (size_t i = 1; i < x_shape.size(); ++i) {
+    result = result * x_shape[i] + pos[i];
   }
-};
+  return result;
+}
 
 template <typename Place, typename T, size_t D>
 void CropGradFunction(const framework::ExecutionContext& context) {
@@ -89,7 +51,7 @@ void CropGradFunction(const framework::ExecutionContext& context) {
   auto d_x_dims = d_x->dims();
   auto d_out_dims = d_out->dims();
 
-  auto offsets = context.op().GetAttr<std::vector<int>>("offsets");
+  auto offsets = context.op().Attr<std::vector<int>>("offsets");
 
   Eigen::array<std::pair<int, int>, D> paddings;
   for (int i = 0; i < d_out_dims.size(); ++i) {
@@ -107,9 +69,9 @@ template <typename Place, typename T>
 class CropGradKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    size_t dim =
+    size_t rank =
         context.Input<Tensor>(framework::GradVarName("Out"))->dims().size();
-    switch (dim) {
+    switch (rank) {
       case 1:
         CropGradFunction<Place, T, 1>(context);
         break;
@@ -129,7 +91,8 @@ class CropGradKernel : public framework::OpKernel {
         CropGradFunction<Place, T, 6>(context);
         break;
       default:
-        LOG(ERROR) << "Only ranks up to 6 supported.";
+        PADDLE_THROW(
+            "CropOp only support tensors with no more than 6 dimensions.");
     }
   }
 };

paddle/operators/gaussian_random_op.cc

@@ -19,12 +19,12 @@ template <typename T>
 class CPUGaussianRandomKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    float mean = context.GetAttr<float>("mean");
-    float std = context.GetAttr<float>("std");
+    float mean = context.Attr<float>("mean");
+    float std = context.Attr<float>("std");
     auto* tensor = context.Output<framework::Tensor>("Out");
     T* data = tensor->mutable_data<T>(context.GetPlace());
 
-    unsigned int seed = static_cast<unsigned int>(context.GetAttr<int>("seed"));
+    unsigned int seed = static_cast<unsigned int>(context.Attr<int>("seed"));
     std::minstd_rand engine;
     if (seed == 0) {
       seed = std::random_device()();
@@ -45,7 +45,7 @@ class GaussianRandomOp : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext& context) const override {
     auto* tensor = context.Output<framework::Tensor>("Out");
-    auto dims = GetAttr<std::vector<int>>("dims");
+    auto dims = Attr<std::vector<int>>("dims");
     std::vector<int64_t> temp;
     temp.reserve(dims.size());
     for (auto dim : dims) {

paddle/operators/gaussian_random_op.cu

@@ -42,13 +42,13 @@ class GPUGaussianRandomKernel : public framework::OpKernel {
   void Compute(const framework::ExecutionContext& context) const override {
     auto* tensor = context.Output<framework::Tensor>("Out");
     T* data = tensor->mutable_data<T>(context.GetPlace());
-    unsigned int seed = static_cast<unsigned int>(context.GetAttr<int>("seed"));
+    unsigned int seed = static_cast<unsigned int>(context.Attr<int>("seed"));
     if (seed == 0) {
       std::random_device rd;
       seed = rd();
     }
-    T mean = static_cast<T>(context.GetAttr<float>("mean"));
-    T std = static_cast<T>(context.GetAttr<float>("std"));
+    T mean = static_cast<T>(context.Attr<float>("mean"));
+    T std = static_cast<T>(context.Attr<float>("std"));
     thrust::counting_iterator<unsigned int> index_sequence_begin(0);
     ssize_t N = framework::product(tensor->dims());
     thrust::transform(index_sequence_begin, index_sequence_begin + N,

paddle/operators/identity_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/operators/net_op.h"
+#include "paddle/operators/scale_op.h"
+
+namespace paddle {
+namespace operators {
+
+// The identity operator is an alias of the scale operator. This is also an
+// example for creating an alias for an existing operator.
+template <typename AttrType>
+class IdentityOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  IdentityOpMaker(framework::OpProto *proto,
+                  framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "The input tensor of identity operator.");
+    AddOutput("Out", "The output tensor of identity operator.");
+    AddComment(R"DOC(
+The identity operator is an alias of the scale operator
+with the attribute scale fixed to 1.0.
+)DOC");
+  }
+};
+
+template <typename AttrType>
+class IdentityOp : public NetOp {
+ public:
+  IdentityOp(const std::string &type, const framework::VariableNameMap &inputs,
+             const framework::VariableNameMap &outputs,
+             const framework::AttributeMap &attrs)
+      : NetOp(type, inputs, outputs, attrs) {
+    AppendOp(framework::OpRegistry::CreateOp(
+        "scale", {{"X", {Input("X")}}}, {{"Out", {Output("Out")}}},
+        {{"scale", static_cast<AttrType>(1)}}));
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_WITHOUT_GRADIENT(identity, ops::IdentityOp<float>,
+                             ops::IdentityOpMaker<float>);

paddle/operators/math/CMakeLists.txt

 
 if(WITH_GPU)
-    nv_library(math_function SRCS math_function.cc math_function.cu DEPS cblas device_context)
+    nv_library(math_function SRCS math_function.cc math_function.cu im2col.cc 
+    im2col.cu DEPS cblas device_context)
 else()
-    cc_library(math_function SRCS math_function.cc DEPS cblas device_context)
+    cc_library(math_function SRCS math_function.cc im2col.cc DEPS cblas device_context)
 endif()
 
 nv_test(math_function_test SRCS math_function_test.cc DEPS math_function tensor)
+cc_test(im2col_test SRCS im2col_test.cc DEPS math_function tensor)

paddle/operators/rnn/recurrent_op_utils.cc

@@ -109,7 +109,7 @@ void InitArgument(const ArgumentName& name, Argument* arg,
   arg->step_scopes = op.Output(name.step_scopes);
 
   auto inlinks = op.Inputs(name.inlinks);
-  auto inlink_alias = op.GetAttr<std::vector<std::string>>(name.inlink_alias);
+  auto inlink_alias = op.Attr<std::vector<std::string>>(name.inlink_alias);
   PADDLE_ENFORCE(inlinks.size() == inlink_alias.size(),
                  "the size of inlinks and inlink_alias don't match:%d,%d",
                  inlinks.size(), inlink_alias.size());
@@ -121,7 +121,7 @@ void InitArgument(const ArgumentName& name, Argument* arg,
   }
 
   auto outlinks = op.Outputs(name.outlinks);
-  auto outlink_alias = op.GetAttr<std::vector<std::string>>(name.outlink_alias);
+  auto outlink_alias = op.Attr<std::vector<std::string>>(name.outlink_alias);
   PADDLE_ENFORCE(outlinks.size() == outlink_alias.size(),
                  "the size of outlinks and outlink_alias don't match:%d,%d",
                  outlinks.size(), outlink_alias.size());
@@ -135,8 +135,8 @@ void InitArgument(const ArgumentName& name, Argument* arg,
   auto boot_memories = op.Inputs(name.boot_memories);
 
   // attributes
-  auto memories = op.GetAttr<std::vector<std::string>>(name.memories);
-  auto pre_memories = op.GetAttr<std::vector<std::string>>(name.pre_memories);
+  auto memories = op.Attr<std::vector<std::string>>(name.memories);
+  auto pre_memories = op.Attr<std::vector<std::string>>(name.pre_memories);
 
   PADDLE_ENFORCE(memories.size() == boot_memories.size(),
                  "the size of memories, boot_memories don't match:%d,%d",

paddle/operators/rowwise_add_op.h

@@ -33,10 +33,12 @@ class RowwiseAddKernel : public framework::OpKernel {
   void Compute(const framework::ExecutionContext& context) const override {
     auto out = context.Output<Tensor>("Out");
     out->mutable_data<T>(context.GetPlace());
-
-    auto input = EigenMatrix<T>::From(*context.Input<Tensor>("X"));
-    auto bias = EigenVector<T>::From(*context.Input<Tensor>("b"));
-    auto output = EigenMatrix<T>::From(*out);
+    int num_col_dims = context.Input<Tensor>("X")->dims().size() -
+                       context.Input<Tensor>("b")->dims().size();
+    auto input =
+        EigenMatrix<T>::Reshape(*context.Input<Tensor>("X"), num_col_dims);
+    auto bias = EigenVector<T>::Flatten(*context.Input<Tensor>("b"));
+    auto output = EigenMatrix<T>::Reshape(*out, num_col_dims);
 
     const int bias_size = bias.dimension(0);
     const int rest_size = input.size() / bias_size;
@@ -54,12 +56,15 @@ class RowwiseAddGradKernel : public framework::OpKernel {
     auto* dout = context.Input<Tensor>(framework::GradVarName("Out"));
     auto* dx = context.Output<Tensor>(framework::GradVarName("X"));
     auto* db = context.Output<Tensor>(framework::GradVarName("b"));
+    int num_col_dims = context.Input<Tensor>("X")->dims().size() -
+                       context.Input<Tensor>("b")->dims().size();
 
-    auto out_grad = EigenMatrix<T>::From(*dout);
+    auto out_grad = EigenMatrix<T>::Reshape(*dout, num_col_dims);
     auto place = context.GetEigenDevice<Place>();
+
     if (dx) {
       dx->mutable_data<T>(context.GetPlace());
-      EigenMatrix<T>::From(*dx).device(place) = out_grad;
+      EigenMatrix<T>::Reshape(*dx, num_col_dims).device(place) = out_grad;
     }
 
     if (db) {