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

.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:

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/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/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/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 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();
+  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 h = inputLayers_[0]->getOutput().getFrameHeight();
+  if (h != 0) inDims_.setDim(2, h);
+  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

@@ -2008,6 +2008,31 @@ 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

@@ -59,7 +59,7 @@ set(DEPS_OPS
 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 
+op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc
   DEPS framework_proto tensor operator net_op)
 op_library(scale_op DEPS net_op)
 

paddle/operators/identity_op.cc

@@ -18,17 +18,20 @@
 namespace paddle {
 namespace operators {
 
-// identity is a alias of scale op. This is also a example for creating a alias
-// operator.
+// 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", "input tensor of identity op");
-    AddOutput("Out", "output tensor of identity op");
-    AddComment("identity operator. Just a alias of scale op which scale = 1.0");
+    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");
   }
 };
 

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/math/im2col.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/math/im2col.h"
+
+namespace paddle {
+namespace operators {
+namespace math {
+
+/*
+ * im = [input_channels, input_height, input_width]
+ * col =
+ *   [input_channels, filter_height, filter_width, output_height, output_width]
+ */
+template <class T>
+class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
+                    platform::CPUPlace, T> {
+ public:
+  void operator()(const framework::Tensor& im, framework::Tensor& col,
+                  int stride_height, int stride_width, int padding_height,
+                  int padding_width, platform::DeviceContext* context) {
+    PADDLE_ENFORCE(im.dims().size() == 3);
+    PADDLE_ENFORCE(col.dims().size() == 5);
+
+    int input_channels = im.dims()[0];
+    int input_height = im.dims()[1];
+    int input_width = im.dims()[2];
+    int filter_height = col.dims()[1];
+    int filter_width = col.dims()[2];
+    int output_height = col.dims()[3];
+    int output_width = col.dims()[4];
+    int channels_col = input_channels * filter_height * filter_width;
+
+    const T* im_data = im.data<T>();
+    T* col_data = col.data<T>();
+
+    for (int c = 0; c < channels_col; ++c) {
+      int w_offset = c % filter_width;
+      int h_offset = (c / filter_width) % filter_height;
+      int c_im = c / filter_width / filter_height;
+      for (int h = 0; h < output_height; ++h) {
+        for (int w = 0; w < output_width; ++w) {
+          int im_row_idx = h * stride_height + h_offset;
+          int im_col_idx = w * stride_width + w_offset;
+          if ((im_row_idx - padding_height) < 0 ||
+              (im_row_idx - padding_height) >= input_height ||
+              (im_col_idx - padding_width) < 0 ||
+              (im_col_idx - padding_width) >= input_width) {
+            col_data[(c * output_height + h) * output_width + w] = T(0);
+          } else {
+            im_row_idx += c_im * input_height - padding_height;
+            im_col_idx -= padding_width;
+            col_data[(c * output_height + h) * output_width + w] =
+                im_data[im_row_idx * input_width + im_col_idx];
+          }
+        }
+      }
+    }
+  }
+};
+
+/*
+ * im = [input_channels, input_height, input_width]
+ * col =
+ *   [input_channels, filter_height, filter_width, output_height, output_width]
+ */
+template <class T>
+class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
+                    platform::CPUPlace, T> {
+ public:
+  void operator()(framework::Tensor& im, const framework::Tensor& col,
+                  int stride_height, int stride_width, int padding_height,
+                  int padding_width, platform::DeviceContext* context) {
+    PADDLE_ENFORCE(im.dims().size() == 3);
+    PADDLE_ENFORCE(col.dims().size() == 5);
+    int input_channels = im.dims()[0];
+    int input_height = im.dims()[1];
+    int input_width = im.dims()[2];
+    int filter_height = col.dims()[1];
+    int filter_width = col.dims()[2];
+    int output_height = col.dims()[3];
+    int output_width = col.dims()[4];
+    int channels_col = input_channels * filter_height * filter_width;
+
+    T* im_data = im.data<T>();
+    const T* col_data = col.data<T>();
+
+    for (int c = 0; c < channels_col; ++c) {
+      int w_offset = c % filter_width;
+      int h_offset = (c / filter_width) % filter_height;
+      int c_im = c / filter_width / filter_height;
+      for (int h = 0; h < output_height; ++h) {
+        for (int w = 0; w < output_width; ++w) {
+          int im_row_idx = h * stride_height + h_offset;
+          int im_col_idx = w * stride_width + w_offset;
+          if ((im_row_idx - padding_height) >= 0 &&
+              (im_row_idx - padding_height) < input_height &&
+              (im_col_idx - padding_width) >= 0 &&
+              (im_col_idx - padding_width) < input_width) {
+            im_row_idx += c_im * input_height - padding_height;
+            im_col_idx -= padding_width;
+            im_data[im_row_idx * input_width + im_col_idx] +=
+                col_data[(c * output_height + h) * output_width + w];
+          }
+        }
+      }
+    }
+  }
+};
+
+template class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
+                             platform::CPUPlace, float>;
+template class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
+                             platform::CPUPlace, double>;
+template class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
+                             platform::CPUPlace, float>;
+template class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
+                             platform::CPUPlace, double>;
+
+/*
+ * im = [input_channels, input_height, input_width]
+ * col =
+ *   [output_height, output_width, input_channels, filter_height, filter_width]
+ */
+template <class T>
+class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
+                    platform::CPUPlace, T> {
+ public:
+  void operator()(const framework::Tensor& im, framework::Tensor& col,
+                  int stride_height, int stride_width, int padding_height,
+                  int padding_width, platform::DeviceContext* context) {
+    PADDLE_ENFORCE(im.dims().size() == 3);
+    PADDLE_ENFORCE(col.dims().size() == 5);
+    int input_channels = im.dims()[0];
+    int input_height = im.dims()[1];
+    int input_width = im.dims()[2];
+    int filter_height = col.dims()[3];
+    int filter_width = col.dims()[4];
+    int output_height = col.dims()[0];
+    int output_width = col.dims()[1];
+
+    const T* im_data = im.data<T>();
+    T* col_data = col.data<T>();
+
+    for (int col_row_idx = 0; col_row_idx < output_height; ++col_row_idx) {
+      for (int col_col_idx = 0; col_col_idx < output_width; ++col_col_idx) {
+        for (int channel = 0; channel < input_channels; ++channel) {
+          for (int filter_row_idx = 0; filter_row_idx < filter_height;
+               ++filter_row_idx) {
+            for (int filter_col_idx = 0; filter_col_idx < filter_width;
+                 ++filter_col_idx) {
+              int im_row_offset =
+                  col_row_idx * stride_height + filter_row_idx - padding_height;
+              int im_col_offset =
+                  col_col_idx * stride_width + filter_col_idx - padding_width;
+              int col_offset = (((col_row_idx * output_width + col_col_idx) *
+                                     input_channels +
+                                 channel) *
+                                    filter_height +
+                                filter_row_idx) *
+                                   filter_width +
+                               filter_col_idx;
+              if (im_row_offset < 0 || im_row_offset >= input_height ||
+                  im_col_offset < 0 || im_col_offset >= input_width) {
+                col_data[col_offset] = T(0);
+              } else {
+                int im_offset =
+                    (channel * input_height + im_row_offset) * input_width +
+                    im_col_offset;
+                col_data[col_offset] = im_data[im_offset];
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+};
+
+/*
+ * im = [input_channels, input_height, input_width]
+ * col =
+ *   [output_height, output_width, input_channels, filter_height, filter_width]
+ */
+template <class T>
+class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
+                    platform::CPUPlace, T> {
+ public:
+  void operator()(framework::Tensor& im, const framework::Tensor& col,
+                  int stride_height, int stride_width, int padding_height,
+                  int padding_width, platform::DeviceContext* context) {
+    PADDLE_ENFORCE(im.dims().size() == 3);
+    PADDLE_ENFORCE(col.dims().size() == 5);
+    int input_channels = im.dims()[0];
+    int input_height = im.dims()[1];
+    int input_width = im.dims()[2];
+    int filter_height = col.dims()[3];
+    int filter_width = col.dims()[4];
+    int output_height = col.dims()[0];
+    int output_width = col.dims()[1];
+
+    T* im_data = im.data<T>();
+    const T* col_data = col.data<T>();
+
+    for (int col_row_idx = 0; col_row_idx < output_height; ++col_row_idx) {
+      for (int col_col_idx = 0; col_col_idx < output_width; ++col_col_idx) {
+        for (int channel = 0; channel < input_channels; ++channel) {
+          for (int filter_row_idx = 0; filter_row_idx < filter_height;
+               ++filter_row_idx) {
+            for (int filter_col_idx = 0; filter_col_idx < filter_width;
+                 ++filter_col_idx) {
+              int im_row_offset =
+                  col_row_idx * stride_height + filter_row_idx - padding_height;
+              int im_col_offset =
+                  col_col_idx * stride_width + filter_col_idx - padding_width;
+              int col_offset = (((col_row_idx * output_width + col_col_idx) *
+                                     input_channels +
+                                 channel) *
+                                    filter_height +
+                                filter_row_idx) *
+                                   filter_width +
+                               filter_col_idx;
+              if (im_row_offset >= 0 && im_row_offset < input_height &&
+                  im_col_offset >= 0 && im_col_offset < input_width) {
+                int im_offset =
+                    (channel * input_height + im_row_offset) * input_width +
+                    im_col_offset;
+                im_data[im_offset] += col_data[col_offset];
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+};
+
+template class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
+                             platform::CPUPlace, float>;
+template class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
+                             platform::CPUPlace, double>;
+template class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
+                             platform::CPUPlace, float>;
+template class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
+                             platform::CPUPlace, double>;
+
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/operators/math/im2col.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/math/im2col.h"
+#include "paddle/platform/cuda_helper.h"
+
+namespace paddle {
+namespace operators {
+namespace math {
+
+template <class T>
+__global__ void im2col(const T* data_im, int num_outs, int height, int width,
+                       int filter_height, int filter_width, int stride_height,
+                       int stride_width, int padding_height, int padding_width,
+                       int output_height, int output_width, T* data_col) {
+  int index = (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x;
+  if (index < num_outs) {
+    int w_out = index % output_width;
+    index /= output_width;
+    int h_out = index % output_height;
+    int channel_in = index / output_height;
+    int channel_out = channel_in * filter_height * filter_width;
+    int h_in = h_out * stride_height;
+    int w_in = w_out * stride_width;
+
+    data_col += (channel_out * output_height + h_out) * output_width + w_out;
+    for (int i = 0; i < filter_height; ++i) {
+      for (int j = 0; j < filter_width; ++j) {
+        int rIdx = int(h_in + i);
+        int cIdx = int(w_in + j);
+        if ((rIdx - (int)padding_height) >= (int)height ||
+            (rIdx - (int)padding_height) < 0 ||
+            (cIdx - (int)padding_width) >= (int)width ||
+            (cIdx - (int)padding_width) < 0) {
+          *data_col = 0;
+        } else {
+          rIdx = rIdx + channel_in * height - padding_height;
+          cIdx = cIdx - padding_width;
+          *data_col = data_im[rIdx * width + cIdx];
+        }
+        data_col += output_height * output_width;
+      }
+    }
+  }
+}
+
+/*
+ * im = [input_channels, input_height, input_width]
+ * col =
+ *   [input_channels, filter_height, filter_width, output_height, output_width]
+ */
+template <class T>
+class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
+                    platform::GPUPlace, T> {
+ public:
+  void operator()(const framework::Tensor& im, framework::Tensor& col,
+                  int stride_height, int stride_width, int padding_height,
+                  int padding_width, platform::DeviceContext* context) {
+    PADDLE_ENFORCE(im.dims().size() == 3);
+    PADDLE_ENFORCE(col.dims().size() == 5);
+
+    int input_channels = im.dims()[0];
+    int input_height = im.dims()[1];
+    int input_width = im.dims()[2];
+    int filter_height = col.dims()[1];
+    int filter_width = col.dims()[2];
+    int output_height = col.dims()[3];
+    int output_width = col.dims()[4];
+
+    int num_outputs = input_channels * output_height * output_width;
+    int blocks = (num_outputs + 1024 - 1) / 1024;
+    int block_x = 512;
+    int block_y = (blocks + 512 - 1) / 512;
+    dim3 threads(1024, 1);
+    dim3 grid(block_x, block_y);
+    im2col<T><<<
+        grid, threads, 0,
+        reinterpret_cast<platform::CUDADeviceContext*>(context)->stream()>>>(
+        im.data<T>(), num_outputs, input_height, input_width, filter_height,
+        filter_width, stride_height, stride_width, padding_height,
+        padding_width, output_height, output_width, col.data<T>());
+  }
+};
+
+template <class T>
+__global__ void col2im(size_t n, const T* data_col, size_t height, size_t width,
+                       size_t channels, size_t filter_height,
+                       size_t filter_width, size_t stride_height,
+                       size_t stride_width, size_t padding_height,
+                       size_t padding_width, size_t output_height,
+                       size_t output_width, T* data_im) {
+  size_t index =
+      (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x;
+  if (index < n) {
+    T val = 0;
+    int w = int(index % width);
+    int h = int((index / width) % height);
+    int c = int(index / (width * height));
+    if ((w - (int)padding_width) >= 0 &&
+        (w - (int)padding_width) < (width - 2 * padding_width) &&
+        (h - (int)padding_height) >= 0 &&
+        (h - padding_height) < (height - 2 * padding_height)) {
+      // compute the start and end of the output
+      int w_col_start = (w < (int)filter_width)
+                            ? 0
+                            : (w - int(filter_width)) / (int)stride_width + 1;
+      int w_col_end =
+          min((int)(w / (int)stride_width + 1), (int)(output_width));
+      int h_col_start = (h < (int)filter_height)
+                            ? 0
+                            : (h - (int)filter_height) / (int)stride_height + 1;
+      int h_col_end = min(int(h / stride_height + 1), int(output_height));
+      for (int h_col = h_col_start; h_col < h_col_end; ++h_col) {
+        for (int w_col = w_col_start; w_col < w_col_end; ++w_col) {
+          // the col location: [c * width * height + h_out, w_out]
+          int c_col = int(c * filter_height * filter_width) +
+                      (h - h_col * (int)stride_height) * (int)filter_width +
+                      (w - w_col * (int)stride_width);
+          val +=
+              data_col[(c_col * output_height + h_col) * output_width + w_col];
+        }
+      }
+      h -= padding_height;
+      w -= padding_width;
+      data_im[c * ((width - 2 * padding_width) *
+                   (height - 2 * padding_height)) +
+              h * (width - 2 * padding_width) + w] += val;
+    }
+  }
+}
+
+/*
+ * im = [input_channels, input_height, input_width]
+ * col =
+ *   [input_channels, filter_height, filter_width, output_height, output_width]
+ */
+template <class T>
+class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
+                    platform::GPUPlace, T> {
+ public:
+  void operator()(framework::Tensor& im, const framework::Tensor& col,
+                  int stride_height, int stride_width, int padding_height,
+                  int padding_width, platform::DeviceContext* context) {
+    PADDLE_ENFORCE(im.dims().size() == 3);
+    PADDLE_ENFORCE(col.dims().size() == 5);
+
+    int input_channels = im.dims()[0];
+    int input_height = im.dims()[1];
+    int input_width = im.dims()[2];
+    int filter_height = col.dims()[1];
+    int filter_width = col.dims()[2];
+    int output_height = col.dims()[3];
+    int output_width = col.dims()[4];
+
+    size_t num_kernels = input_channels * (input_height + 2 * padding_height) *
+                         (input_width + 2 * padding_width);
+
+    size_t blocks = (num_kernels + 1024 - 1) / 1024;
+    size_t block_x = 512;
+    size_t block_y = (blocks + 512 - 1) / 512;
+    dim3 threads(1024, 1);
+    dim3 grid(block_x, block_y);
+
+    // To avoid involving atomic operations, we will launch one kernel per
+    // bottom dimension, and then in the kernel add up the top dimensions.
+    col2im<T><<<
+        grid, threads, 0,
+        reinterpret_cast<platform::CUDADeviceContext*>(context)->stream()>>>(
+        num_kernels, col.data<T>(), input_height + 2 * padding_height,
+        input_width + 2 * padding_width, input_channels, filter_height,
+        filter_width, stride_height, stride_width, padding_height,
+        padding_width, output_height, output_width, im.data<T>());
+  }
+};
+
+template class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
+                             platform::GPUPlace, float>;
+template class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
+                             platform::GPUPlace, double>;
+template class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
+                             platform::GPUPlace, float>;
+template class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
+                             platform::GPUPlace, double>;
+
+template <class T>
+__global__ void im2colOCF(const T* im_data, T* col_data, int input_channels,
+                          int input_height, int input_width, int filter_height,
+                          int filter_width, int stride_height, int stride_width,
+                          int padding_height, int padding_width,
+                          int output_height, int output_width) {
+  int swid = blockIdx.x;
+  int shid = blockIdx.y;
+  for (int channelid = threadIdx.z; channelid < input_channels;
+       channelid += blockDim.z) {
+    for (int idy = threadIdx.y; idy < filter_height; idy += blockDim.y) {
+      for (int idx = threadIdx.x; idx < filter_width; idx += blockDim.x) {
+        int width_offset = idx + swid * stride_width - padding_width;
+        int height_offset = idy + shid * stride_height - padding_height;
+        int im_offset = width_offset + height_offset * input_width +
+                        channelid * input_height * input_width;
+
+        int col_offset = idx + idy * filter_width +
+                         channelid * filter_height * filter_width +
+                         (shid * output_width + swid) *
+                             (input_channels * filter_height * filter_width);
+
+        if (height_offset >= input_height || height_offset < 0 ||
+            width_offset >= input_width || width_offset < 0) {
+          col_data[col_offset] = T(0);
+        } else {
+          col_data[col_offset] = im_data[im_offset];
+        }
+      }
+    }
+  }
+}
+
+/*
+ * im = [input_channels, input_height, input_width]
+ * col =
+ *   [output_height, output_width, input_channels, filter_height, filter_width]
+ */
+template <class T>
+class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
+                    platform::GPUPlace, T> {
+ public:
+  void operator()(const framework::Tensor& im, framework::Tensor& col,
+                  int stride_height, int stride_width, int padding_height,
+                  int padding_width, platform::DeviceContext* context) {
+    PADDLE_ENFORCE(im.dims().size() == 3);
+    PADDLE_ENFORCE(col.dims().size() == 5);
+    int input_channels = im.dims()[0];
+    int input_height = im.dims()[1];
+    int input_width = im.dims()[2];
+    int filter_height = col.dims()[3];
+    int filter_width = col.dims()[4];
+    int output_height = col.dims()[0];
+    int output_width = col.dims()[1];
+
+    int block_dim_x = 0;
+    int block_dim_y = 0;
+    if (filter_height <= 4 && filter_width <= 4) {
+      block_dim_x = 4;
+      block_dim_y = 4;
+    } else if (filter_height <= 8 && filter_width <= 8) {
+      block_dim_x = 8;
+      block_dim_y = 8;
+    } else if (filter_height <= 16 && filter_width <= 16) {
+      block_dim_x = 16;
+      block_dim_y = 16;
+    } else {
+      block_dim_x = 32;
+      block_dim_y = 32;
+    }
+
+    int block_dim_z = 1024 / block_dim_x / block_dim_y;
+    dim3 threads(block_dim_x, block_dim_y,
+                 std::min(block_dim_z, input_channels));
+    dim3 grid(output_width, output_height);
+    im2colOCF<T><<<
+        grid, threads, 0,
+        reinterpret_cast<platform::CUDADeviceContext*>(context)->stream()>>>(
+        im.data<T>(), col.data<T>(), input_channels, input_height, input_width,
+        filter_height, filter_width, stride_height, stride_width,
+        padding_height, padding_width, output_height, output_width);
+  }
+};
+
+template <class T>
+__global__ void col2imOCF(T* im_data, const T* col_data, int input_channels,
+                          int input_height, int input_width, int filter_height,
+                          int filter_width, int stride_height, int stride_width,
+                          int padding_height, int padding_width,
+                          int output_height, int output_width) {
+  int swid = blockIdx.x;
+  int shid = blockIdx.y;
+  for (int channelid = threadIdx.z; channelid < input_channels;
+       channelid += blockDim.z) {
+    for (int idy = threadIdx.y; idy < filter_height; idy += blockDim.y) {
+      for (int idx = threadIdx.x; idx < filter_width; idx += blockDim.x) {
+        int width_offset = idx + swid * stride_width - padding_width;
+        int height_offset = idy + shid * stride_height - padding_height;
+        int im_offset = width_offset + height_offset * input_width +
+                        channelid * input_height * input_width;
+
+        int col_offset = idx + idy * filter_width +
+                         channelid * filter_height * filter_width +
+                         (shid * output_width + swid) *
+                             (input_channels * filter_height * filter_width);
+
+        if (height_offset >= 0 && height_offset < input_height &&
+            width_offset >= 0 && width_offset < input_width) {
+          paddle::platform::CudaAtomicAdd(im_data + im_offset,
+                                          col_data[col_offset]);
+        }
+      }
+    }
+  }
+}
+
+/*
+ * im = [input_channels, input_height, input_width]
+ * col =
+ *   [output_height, output_width, input_channels, filter_height, filter_width]
+ */
+template <class T>
+class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
+                    platform::GPUPlace, T> {
+ public:
+  void operator()(framework::Tensor& im, const framework::Tensor& col,
+                  int stride_height, int stride_width, int padding_height,
+                  int padding_width, platform::DeviceContext* context) {
+    PADDLE_ENFORCE(im.dims().size() == 3);
+    PADDLE_ENFORCE(col.dims().size() == 5);
+    int input_channels = im.dims()[0];
+    int input_height = im.dims()[1];
+    int input_width = im.dims()[2];
+    int filter_height = col.dims()[3];
+    int filter_width = col.dims()[4];
+    int output_height = col.dims()[0];
+    int output_width = col.dims()[1];
+
+    int block_dim_x = 0;
+    int block_dim_y = 0;
+    if (filter_height <= 4 && filter_width <= 4) {
+      block_dim_x = 4;
+      block_dim_y = 4;
+    } else if (filter_height <= 8 && filter_width <= 8) {
+      block_dim_x = 8;
+      block_dim_y = 8;
+    } else if (filter_height <= 16 && filter_width <= 16) {
+      block_dim_x = 16;
+      block_dim_y = 16;
+    } else {
+      block_dim_x = 32;
+      block_dim_y = 32;
+    }
+
+    int block_dim_z = 1024 / block_dim_x / block_dim_y;
+    dim3 threads(block_dim_x, block_dim_y,
+                 std::min(block_dim_z, input_channels));
+    dim3 grid(output_width, output_height);
+    col2imOCF<T><<<
+        grid, threads, 0,
+        reinterpret_cast<platform::CUDADeviceContext*>(context)->stream()>>>(
+        im.data<T>(), col.data<T>(), input_channels, input_height, input_width,
+        filter_height, filter_width, stride_height, stride_width,
+        padding_height, padding_width, output_height, output_width);
+  }
+};
+
+template class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
+                             platform::GPUPlace, float>;
+template class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
+                             platform::GPUPlace, double>;
+template class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
+                             platform::GPUPlace, float>;
+template class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
+                             platform::GPUPlace, double>;
+
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/operators/math/im2col.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 "paddle/framework/tensor.h"
+#include "paddle/platform/device_context.h"
+
+namespace paddle {
+namespace operators {
+namespace math {
+
+/* The storage format of the coldata in the Im2ColFunctor and Col2ImFunctor. */
+enum class ColFormat { kCFO = 0, kOCF = 1 };
+
+/*
+ * \brief Converts the image data of three dimensions(CHW) into a colData of
+ *        five dimensions in the Im2ColFunctor calculation,
+ *        And in the Col2ImFunctor calculation, it is reversed.
+ *
+ * \param imData   Image data.
+ * \param imShape  The shape of imData,
+ *                 [input_channels, input_height, input_width].
+ * \param colData  Column data.
+ * \param colShape The shape of colData.
+ *
+ * If the template argument Format is kCFO, the shape of colData is:
+ * [input_channels, filter_height, filter_width, output_height, output_width]
+ * So, it is easy to reshape into a convolution matrix for convolution
+ * calculation based on matrix multiplication.
+ * The shape of convolution matrix is [height, width], where the height is equal
+ * input_channels * filter_height * filter_width, and the width is equal
+ * output_height * output_width.
+ *
+ * Reshape:
+ *     shape of colData           shape of convolution matrix
+ *     [input_channels,
+ *      filter_height,
+ *      filter_width,      ======>      [height, width]
+ *      output_height,
+ *      output_width]
+ *
+ * If the template argument Format is kOCF, the shape of colData is:
+ * [output_height, output_width, input_channels, filter_height, filter_width]
+ * So, it is easy to reshape into a sequence matrix for rnn calculation.
+ * The shape of sequence matrix is [seq_length, step_size], where the seq_length
+ * is equal output_height * output_width, and the step_size is equal
+ * input_channels * filter_height * filter_width.
+ *
+ * Reshape:
+ *     shape of colData             shape of sequence matrix
+ *     [output_height,
+ *      output_width,
+ *      input_channels,    ======>    [seqLength, stepSize]
+ *      filter_height,
+ *      filter_width]
+ *
+ * \note The caller needs to ensure that imShape.inputChannels is equal to
+ *       colShape.inputChannels.
+ */
+template <ColFormat Format, typename Place, typename T>
+class Im2ColFunctor {
+ public:
+  void operator()(const framework::Tensor& im, framework::Tensor& col,
+                  int stride_height, int stride_width, int padding_height,
+                  int padding_width, platform::DeviceContext* context);
+};
+
+template <ColFormat Format, typename Place, typename T>
+class Col2ImFunctor {
+ public:
+  void operator()(framework::Tensor& im, const framework::Tensor& col,
+                  int stride_height, int stride_width, int padding_height,
+                  int padding_width, platform::DeviceContext* context);
+};
+
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/operators/math/im2col_test.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/math/im2col.h"
+#include <gtest/gtest.h>
+#include <iostream>
+
+template <typename Place>
+void testIm2col() {
+  paddle::framework::Tensor input_tmp;
+  paddle::framework::Tensor input;
+  paddle::framework::Tensor output_cfo;
+  paddle::framework::Tensor output_ocf;
+  paddle::framework::Tensor output_tmp;
+
+  /**
+   * input = [0, 1, 2,
+   *          3, 4, 5]
+   *
+   * output_cfo = [0, 1
+   *               1, 2
+   *               3, 4
+   *               4, 5]
+   *
+   * output_ocf = [0, 1, 3, 4
+   *               1, 2, 4, 5]
+   */
+  int input_height = 2;
+  int input_width = 3;
+  int filter_size = 2;
+  int stride = 1;
+  int padding = 0;
+  int output_height = (input_height - filter_size + 2 * padding) / stride + 1;
+  int output_width = (input_width - filter_size + 2 * padding) / stride + 1;
+  float* input_ptr = input_tmp.mutable_data<float>(
+      {1, input_height, input_width}, paddle::platform::CPUPlace());
+  float arr[6] = {0, 1, 2, 3, 4, 5};
+  memcpy(input_ptr, arr, 6 * sizeof(float));
+
+  auto* place = new Place();
+  if (paddle::platform::is_cpu_place(*place)) {
+    input = input_tmp;
+  } else {
+    input.CopyFrom<float>(input_tmp, *place);
+  }
+  output_cfo.mutable_data<float>(
+      {1, filter_size, filter_size, output_height, output_width}, *place);
+  output_ocf.mutable_data<float>(
+      {output_height, output_width, 1, filter_size, filter_size}, *place);
+
+  paddle::operators::math::Im2ColFunctor<
+      paddle::operators::math::ColFormat::kCFO, Place, float>
+      im2col;
+  paddle::operators::math::Im2ColFunctor<
+      paddle::operators::math::ColFormat::kOCF, Place, float>
+      im2col_ocf;
+
+  paddle::platform::DeviceContext* context;
+  if (paddle::platform::is_cpu_place(*place)) {
+    context =
+        new paddle::platform::CPUDeviceContext(paddle::platform::CPUPlace());
+  } else {
+#ifndef PADDLE_ONLY_CPU
+    context =
+        new paddle::platform::CUDADeviceContext(paddle::platform::GPUPlace());
+#else
+    PADDLE_THROW("no GPU support");
+#endif  // PADDLE_ONLY_CPU
+  }
+  im2col(input, output_cfo, stride, stride, padding, padding, context);
+  im2col_ocf(input, output_ocf, stride, stride, padding, padding, context);
+
+  float* out_cfo_ptr;
+  if (paddle::platform::is_cpu_place(*place)) {
+    out_cfo_ptr = output_cfo.data<float>();
+  } else {
+    output_tmp.CopyFrom<float>(output_cfo, paddle::platform::CPUPlace());
+    out_cfo_ptr = output_tmp.data<float>();
+  }
+  EXPECT_EQ(out_cfo_ptr[0], 0);
+  EXPECT_EQ(out_cfo_ptr[1], 1);
+  EXPECT_EQ(out_cfo_ptr[2], 1);
+  EXPECT_EQ(out_cfo_ptr[3], 2);
+  EXPECT_EQ(out_cfo_ptr[4], 3);
+  EXPECT_EQ(out_cfo_ptr[5], 4);
+  EXPECT_EQ(out_cfo_ptr[6], 4);
+  EXPECT_EQ(out_cfo_ptr[7], 5);
+
+  float* out_ocf_ptr;
+  if (paddle::platform::is_cpu_place(*place)) {
+    out_ocf_ptr = output_ocf.data<float>();
+  } else {
+    output_tmp.CopyFrom<float>(output_ocf, paddle::platform::CPUPlace());
+    out_ocf_ptr = output_tmp.data<float>();
+  }
+  EXPECT_EQ(out_ocf_ptr[0], 0);
+  EXPECT_EQ(out_ocf_ptr[1], 1);
+  EXPECT_EQ(out_ocf_ptr[2], 3);
+  EXPECT_EQ(out_ocf_ptr[3], 4);
+  EXPECT_EQ(out_ocf_ptr[4], 1);
+  EXPECT_EQ(out_ocf_ptr[5], 2);
+  EXPECT_EQ(out_ocf_ptr[6], 4);
+  EXPECT_EQ(out_ocf_ptr[7], 5);
+}
+
+TEST(math, im2col) {
+  testIm2col<paddle::platform::CPUPlace>();
+#ifndef PADDLE_ONLY_CPU
+  testIm2col<paddle::platform::GPUPlace>();
+#endif
+}
\ No newline at end of file

paddle/operators/scale_op.cc

@@ -44,11 +44,13 @@ class ScaleOpMaker : public framework::OpProtoAndCheckerMaker {
 
 The equation is: Out = scale*X
 )DOC");
-    AddAttr<AttrType>("scale", "scale of scale operator.").SetDefault(1.0);
+    AddAttr<AttrType>("scale", "The scaling factor of the scale operator.")
+        .SetDefault(1.0);
   }
 };
 
-// Scale Op's gradient is scale op, too.
+// The operator to calculate gradients of a scale operator is just the scale
+// operator itself.
 // Grad(Out=scale(X)) => Grad(X) = scale(Grad(Out))
 template <typename AttrType>
 class ScaleGradOp : public NetOp {

paddle/operators/softmax_op.cc

@@ -51,7 +51,7 @@ the other dimensions in the K-dimensional vector input. Then the ratio of the
 exponential of the given dimension and the sum of exponential values of all
 the other dimensions is the output of the softmax operator.
 
-For each row `i` and each column `j` in X, we have:
+For each row `i` and each column `j` in input X, we have:
     Y[i, j] = exp(X[i, j]) / sum_j(exp(X[i, j]))
 
 )DOC");
@@ -64,14 +64,15 @@ class SoftmaxOpGrad : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(ctx.InputVar("Y") != nullptr, "Input(Y) should not be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should be not null.");
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Y")),
-                            "Input(Y@GRAD) should not be null");
-    PADDLE_ENFORCE(ctx.Input<Tensor>("Y")->dims() ==
-                       ctx.Input<Tensor>(framework::GradVarName("Y"))->dims(),
-                   "the shape of Input(0) and Input(1) should be the same");
+                            "Input(Y@GRAD) should be not null.");
+    PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("Y")->dims(),
+                      ctx.Input<Tensor>(framework::GradVarName("Y"))->dims(),
+                      "Input(Y) and its gradients should have a same shape.");
+
     ctx.Output<Tensor>(framework::GradVarName("X"))
-        ->Resize(ctx.Input<Tensor>("Y")->dims());
+        ->Resize(ctx.Input<Tensor>("X")->dims());
   }
 };
 

paddle/operators/softmax_op.h

@@ -28,12 +28,12 @@ template <typename Place, typename T>
 class SoftmaxKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto input = context.Input<Tensor>("X");
-    auto output = context.Output<Tensor>("Y");
-    output->mutable_data<T>(context.GetPlace());
+    auto X = context.Input<Tensor>("X");
+    auto Y = context.Output<Tensor>("Y");
+    Y->mutable_data<T>(context.GetPlace());
 
-    auto logits = EigenMatrix<T>::From(*input);
-    auto softmax = EigenMatrix<T>::From(*output);
+    auto logits = EigenMatrix<T>::From(*X);
+    auto softmax = EigenMatrix<T>::From(*Y);
 
     const int kBatchDim = 0;
     const int kClassDim = 1;

paddle/operators/squared_l2_distance_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/squared_l2_distance_op.h"
+
+namespace paddle {
+namespace operators {
+
+class SquaredL2DistanceOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext& ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input of SquaredL2DistanceOp "
+                            "must be initialized.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
+                            "Target of SquaredL2DistanceOp "
+                            "must be initialized.");
+
+    auto* x = ctx.Input<Tensor>("X");
+    auto x_dims = x->dims();
+    auto* y = ctx.Input<Tensor>("Y");
+    auto y_dims = y->dims();
+
+    PADDLE_ENFORCE_EQ(framework::arity(x_dims), framework::arity(y_dims),
+                      "Tensor rank of both SquaredL2DistanceOp's "
+                      "inputs must be same.");
+
+    int rank = framework::arity(x_dims);
+    PADDLE_ENFORCE_GE(rank, 2, "Tensor rank should be at least equal to 2.");
+    PADDLE_ENFORCE_EQ(framework::product(x_dims) / x_dims[0],
+                      framework::product(y_dims) / y_dims[0],
+                      "Product of dimensions expcet the first dimension of "
+                      "input and target must be equal.");
+    PADDLE_ENFORCE(y_dims[0] == 1 || y_dims[0] == x_dims[0],
+                   "First dimension of target must be equal to input "
+                   "or to 1.");
+
+    ctx.Output<Tensor>("sub_result")
+        ->Resize({static_cast<int>(x_dims[0]),
+                  static_cast<int>(framework::product(x_dims) / x_dims[0])});
+    ctx.Output<Tensor>("Out")->Resize({x_dims[0], 1});
+  }
+};
+
+class SquaredL2DistanceOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  SquaredL2DistanceOpMaker(framework::OpProto* proto,
+                           framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "Input of SquaredL2DistanceOp.");
+    AddInput("Y", "Target of SquaredL2DistanceOp.");
+    AddOutput("sub_result",
+              "Buffering substraction result which "
+              "will be reused in backward.")
+        .AsIntermediate();
+    AddOutput("Out", "Squared l2 distance between input and target.");
+    AddComment(R"DOC(
+    SquaredL2DistanceOp will cacluate the squared L2 distance for
+    input and target. Number of distance value equals to the
+    first dimension of input. First dimension of target could be equal to
+    input or to 1. If the first dimension of target is 1, SquaredL2DistanceOp
+    will broadcast target's first dimension to input's first dimension.
+    You can decide whether calculate the gradient of input and target.
+    )DOC");
+  }
+};
+
+class SquaredL2DistanceGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext& ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
+                            "Gradient of Out should not be null");
+    auto out_dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
+    auto x_dims = ctx.Input<Tensor>("X")->dims();
+    auto y_dims = ctx.Input<Tensor>("Y")->dims();
+    PADDLE_ENFORCE_EQ(out_dims[0], x_dims[0],
+                      "First dimension of output gradient and "
+                      "input value must be equal.");
+    PADDLE_ENFORCE_EQ(out_dims[1], 1,
+                      "Second dimension of output gradient "
+                      "must be 1.");
+    auto* x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto* y_grad = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    if (x_grad) x_grad->Resize(x_dims);
+    if (y_grad) y_grad->Resize(y_dims);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(squared_l2_distance, ops::SquaredL2DistanceOp,
+            ops::SquaredL2DistanceOpMaker, squared_l2_distance_grad,
+            ops::SquaredL2DistanceGradOp);
+REGISTER_OP_CPU_KERNEL(
+    squared_l2_distance,
+    ops::SquaredL2DistanceKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    squared_l2_distance_grad,
+    ops::SquaredL2DistanceGradKernel<paddle::platform::CPUPlace, float>);

paddle/operators/squared_l2_distance_op.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#define EIGEN_USE_GPU
+
+#include "paddle/operators/squared_l2_distance_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(
+    squared_l2_distance,
+    ops::SquaredL2DistanceKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    squared_l2_distance_grad,
+    ops::SquaredL2DistanceGradKernel<paddle::platform::GPUPlace, float>);

paddle/operators/squared_l2_distance_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
+
+template <typename Place, typename T>
+class SquaredL2DistanceKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* in0 = context.Input<Tensor>("X");
+    auto* in1 = context.Input<Tensor>("Y");
+    auto* out0 = context.Output<Tensor>("sub_result");
+    auto* out1 = context.Output<Tensor>("Out");
+
+    auto in0_dims = in0->dims();
+    auto in1_dims = in1->dims();
+
+    int cols = framework::product(in0_dims) / in0_dims[0];
+    // reduce dimensions except the first
+    auto x =
+        EigenMatrix<T>::From(*in0, framework::make_ddim({in0_dims[0], cols}));
+    auto y =
+        EigenMatrix<T>::From(*in1, framework::make_ddim({in1_dims[0], cols}));
+
+    out0->mutable_data<T>(context.GetPlace());
+    out1->mutable_data<T>(context.GetPlace());
+    auto sub_result = EigenMatrix<T>::From(*out0);
+    auto z = EigenVector<T>::Flatten(*out1);
+
+    auto place = context.GetEigenDevice<Place>();
+    auto x_dims = x.dimensions();
+    auto y_dims = y.dimensions();
+    // buffer the substraction result
+    if (y_dims[0] == 1 && x_dims[0] > y_dims[0]) {
+      sub_result.device(place) =
+          x -
+          y.broadcast(Eigen::array<int, 2>({{static_cast<int>(x_dims[0]), 1}}));
+    } else {
+      sub_result.device(place) = x - y;
+    }
+    auto sub_res_pow2 = sub_result * sub_result;
+    z.device(place) = sub_res_pow2.sum(Eigen::array<int, 1>({{1}}));
+  }
+};
+
+template <typename Place, typename T>
+class SquaredL2DistanceGradKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* in0 = context.Input<Tensor>("sub_result");
+    auto* in1 = context.Input<Tensor>(framework::GradVarName("Out"));
+    auto* x_g = context.Output<Tensor>(framework::GradVarName("X"));
+    auto* y_g = context.Output<Tensor>(framework::GradVarName("Y"));
+
+    auto sub_result = EigenMatrix<T>::From(*in0);
+    auto out_grad = EigenMatrix<T>::From(*in1);
+
+    auto x_dims = x_g->dims();
+    auto y_dims = y_g->dims();
+
+    int cols = framework::product(x_dims) / x_dims[0];
+    // calculate gradient
+    auto grad_mat = 2 *
+                    (out_grad.broadcast(Eigen::array<int, 2>({{1, cols}}))) *
+                    sub_result;
+
+    // propagate back to input
+    auto eigen_place = context.GetEigenDevice<Place>();
+    if (x_g) {
+      x_g->mutable_data<T>(context.GetPlace());
+      // eigen matrix
+      auto x_grad =
+          EigenMatrix<T>::From(*x_g, framework::make_ddim({x_dims[0], cols}));
+      // dimensions are same with subResult
+      x_grad.device(eigen_place) = grad_mat;
+    }
+
+    if (y_g) {
+      y_g->mutable_data<T>(context.GetPlace());
+
+      PADDLE_ENFORCE_GE(sub_result.dimensions()[0], y_dims[0],
+                        "First dimension of gradient must be greater or "
+                        "equal than first dimension of target.");
+
+      if (sub_result.dimensions()[0] == y_dims[0]) {
+        auto y_grad =
+            EigenMatrix<T>::From(*y_g, framework::make_ddim({y_dims[0], cols}));
+        y_grad.device(eigen_place) = -1 * grad_mat;
+      } else {
+        auto col_sum_res = -1 * (grad_mat.sum(Eigen::array<int, 1>({{0}})));
+        auto y_grad = EigenVector<T>::Flatten(*y_g);
+        y_grad.device(eigen_place) = col_sum_res;
+      }
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/top_k_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/top_k_op.h"
+
+namespace paddle {
+namespace operators {
+
+class TopkOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input of TopkOP must be initialized.");
+    auto *input = ctx.Input<framework::Tensor>("X");
+    const int k = static_cast<int>(ctx.Attr<int>("k"));
+
+    PADDLE_ENFORCE_GE(k, 1, "k must >= 1");
+    PADDLE_ENFORCE_GE(input->dims().size(), 1, "input must have >= 1d shape");
+    PADDLE_ENFORCE_GE(input->dims()[input->dims().size() - 1], k,
+                      "input must have >= k columns");
+
+    framework::DDim dims = input->dims();
+    dims[dims.size() - 1] = k;
+    ctx.Output<Tensor>("Out")->Resize(dims);
+    ctx.Output<Tensor>("Indices")->Resize(dims);
+  }
+};
+
+class TopkOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  TopkOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "The input of Topk op");
+    AddOutput("Out", "The output tensor of Topk op");
+    AddOutput("Indices", "The indices of Topk elements of input");
+    AddComment(
+        R"DOC(If the input is a vector (1d tensor), finds the k largest entries in the vector and outputs their values and indices as vectors. Thus values[j] is the j-th largest entry in input, and its index is indices[j].
+
+    For matrices, computes the top k entries in each row. )DOC");
+    AddAttr<int>("k",
+                 "Number of top elements to look for along the last "
+                 "dimension (along each row for matrices).")
+        .SetDefault(1);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(top_k, ops::TopkOp, ops::TopkOpMaker);
+REGISTER_OP_CPU_KERNEL(top_k,
+                       ops::TopkKernel<paddle::platform::CPUPlace, float>);

paddle/operators/top_k_op.cu

+/* Copyright (c) 2016 PaddlePaddle Authors All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/framework/op_registry.h"
+#include "paddle/platform/assert.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename T>
+struct Pair {
+  __device__ __forceinline__ Pair() {}
+  __device__ __forceinline__ Pair(T value, int id) : v(value), id(id) {}
+
+  __device__ __forceinline__ void set(T value, int id) {
+    v = value;
+    id = id;
+  }
+
+  __device__ __forceinline__ void operator=(const Pair<T>& in) {
+    v = in.v;
+    id = in.id;
+  }
+
+  __device__ __forceinline__ bool operator<(const T value) const {
+    return (v < value);
+  }
+
+  __device__ __forceinline__ bool operator<(const Pair<T>& in) const {
+    return (v < in.v) || ((v == in.v) && (id > in.id));
+  }
+
+  __device__ __forceinline__ bool operator>(const Pair<T>& in) const {
+    return (v > in.v) || ((v == in.v) && (id < in.id));
+  }
+
+  T v;
+  int id;
+};
+
+template <typename T>
+__device__ __forceinline__ void AddTo(Pair<T> topk[], const Pair<T>& p,
+                                      int beam_size) {
+  for (int k = beam_size - 2; k >= 0; k--) {
+    if (topk[k] < p) {
+      topk[k + 1] = topk[k];
+    } else {
+      topk[k + 1] = p;
+      return;
+    }
+  }
+  topk[0] = p;
+}
+
+template <typename T, int beam_size>
+__device__ __forceinline__ void AddTo(Pair<T> topk[], const Pair<T>& p) {
+  for (int k = beam_size - 2; k >= 0; k--) {
+    if (topk[k] < p) {
+      topk[k + 1] = topk[k];
+    } else {
+      topk[k + 1] = p;
+      return;
+    }
+  }
+  topk[0] = p;
+}
+
+template <typename T, int BlockSize>
+__device__ __forceinline__ void GetTopK(Pair<T> topk[], const T* src, int idx,
+                                        int dim, int beam_size) {
+  while (idx < dim) {
+    if (topk[beam_size - 1] < src[idx]) {
+      Pair<T> tmp(src[idx], idx);
+      AddTo<T>(topk, tmp, beam_size);
+    }
+    idx += BlockSize;
+  }
+}
+
+template <typename T, int BlockSize>
+__device__ __forceinline__ void GetTopK(Pair<T> topk[], const T* src, int idx,
+                                        int dim, const Pair<T>& max,
+                                        int beam_size) {
+  while (idx < dim) {
+    if (topk[beam_size - 1] < src[idx]) {
+      Pair<T> tmp(src[idx], idx);
+      if (tmp < max) {
+        AddTo<T>(topk, tmp, beam_size);
+      }
+    }
+    idx += BlockSize;
+  }
+}
+
+template <typename T, int BlockSize>
+__device__ __forceinline__ void GetTopK(Pair<T> topk[], const T* val, int* col,
+                                        int idx, int dim, int beam_size) {
+  while (idx < dim) {
+    if (topk[beam_size - 1] < val[idx]) {
+      Pair<T> tmp(val[idx], col[idx]);
+      AddTo<T>(topk, tmp, beam_size);
+    }
+    idx += BlockSize;
+  }
+}
+
+template <typename T, int BlockSize>
+__device__ __forceinline__ void GetTopK(Pair<T> topk[], const T* val, int* col,
+                                        int idx, int dim, const Pair<T>& max,
+                                        int beam_size) {
+  while (idx < dim) {
+    if (topk[beam_size - 1] < val[idx]) {
+      Pair<T> tmp(val[idx], col[idx]);
+      if (tmp < max) {
+        AddTo<T>(topk, tmp, beam_size);
+      }
+    }
+    idx += BlockSize;
+  }
+}
+
+template <typename T, int MaxLength, int BlockSize>
+__device__ __forceinline__ void ThreadGetTopK(Pair<T> topk[], int& beam,
+                                              int beam_size, const T* src,
+                                              bool& firstStep, bool& is_empty,
+                                              Pair<T>& max, int dim,
+                                              const int tid) {
+  if (beam > 0) {
+    int length = beam < beam_size ? beam : beam_size;
+    if (firstStep) {
+      firstStep = false;
+      GetTopK<T, BlockSize>(topk, src, tid, dim, length);
+    } else {
+      for (int k = 0; k < MaxLength; k++) {
+        if (k < MaxLength - beam) {
+          topk[k] = topk[k + beam];
+        } else {
+          topk[k].set(-INFINITY, -1);
+        }
+      }
+      if (!is_empty) {
+        GetTopK<T, BlockSize>(topk + MaxLength - beam, src, tid, dim, max,
+                              length);
+      }
+    }
+
+    max = topk[MaxLength - 1];
+    if (max.v == -1) is_empty = true;
+    beam = 0;
+  }
+}
+
+template <typename T, int MaxLength, int BlockSize>
+__device__ __forceinline__ void ThreadGetTopK(Pair<T> topk[], int& beam,
+                                              int beam_size, const T* val,
+                                              int* col, bool& firstStep,
+                                              bool& is_empty, Pair<T>& max,
+                                              int dim, const int tid) {
+  if (beam > 0) {
+    int length = beam < beam_size ? beam : beam_size;
+    if (firstStep) {
+      firstStep = false;
+      GetTopK<T, BlockSize>(topk, val, col, tid, dim, length);
+    } else {
+      for (int k = 0; k < MaxLength; k++) {
+        if (k < MaxLength - beam) {
+          topk[k] = topk[k + beam];
+        } else {
+          topk[k].set(-INFINITY, -1);
+        }
+      }
+      if (!is_empty) {
+        GetTopK<T, BlockSize>(topk + MaxLength - beam, val, col, tid, dim, max,
+                              length);
+      }
+    }
+
+    max = topk[MaxLength - 1];
+    if (max.v == -1) is_empty = true;
+    beam = 0;
+  }
+}
+
+template <typename T, int MaxLength, int BlockSize>
+__device__ __forceinline__ void BlockReduce(Pair<T>* sh_topk, int* maxid,
+                                            Pair<T> topk[], T** topVal,
+                                            int** topIds, int& beam, int& k,
+                                            const int tid, const int warp) {
+  while (true) {
+    __syncthreads();
+    if (tid < BlockSize / 2) {
+      if (sh_topk[tid] < sh_topk[tid + BlockSize / 2]) {
+        maxid[tid] = tid + BlockSize / 2;
+      } else {
+        maxid[tid] = tid;
+      }
+    }
+    __syncthreads();
+    for (int stride = BlockSize / 4; stride > 0; stride = stride / 2) {
+      if (tid < stride) {
+        if (sh_topk[maxid[tid]] < sh_topk[maxid[tid + stride]]) {
+          maxid[tid] = maxid[tid + stride];
+        }
+      }
+      __syncthreads();
+    }
+    __syncthreads();
+
+    if (tid == 0) {
+      **topVal = sh_topk[maxid[0]].v;
+      **topIds = sh_topk[maxid[0]].id;
+      (*topVal)++;
+      (*topIds)++;
+    }
+    if (tid == maxid[0]) beam++;
+    if (--k == 0) break;
+    __syncthreads();
+
+    if (tid == maxid[0]) {
+      if (beam < MaxLength) {
+        sh_topk[tid] = topk[beam];
+      }
+    }
+    if (maxid[0] / 32 == warp) {
+      if (__shfl(beam, (maxid[0]) % 32, 32) == MaxLength) break;
+    }
+  }
+}
+
+/**
+ * Each block compute one sample.
+ * In a block:
+ * 1. every thread get top MaxLength value;
+ * 2. merge to sh_topk, block reduce and get max value;
+ * 3. go to the second setp, until one thread's topk value is null;
+ * 4. go to the first setp, until get the topk value.
+ */
+template <typename T, int MaxLength, int BlockSize>
+__global__ void KeMatrixTopK(T* output, int output_stride, int* indices,
+                             const T* src, int lds, int dim, int k) {
+  __shared__ Pair<T> sh_topk[BlockSize];
+  __shared__ int maxid[BlockSize / 2];
+  const int tid = threadIdx.x;
+  const int warp = threadIdx.x / 32;
+  output += blockIdx.x * output_stride;
+  indices += blockIdx.x * k;
+
+  Pair<T> topk[MaxLength];
+  int beam = MaxLength;
+  Pair<T> max;
+  bool is_empty = false;
+  bool firststep = true;
+
+  for (int k = 0; k < MaxLength; k++) {
+    topk[k].set(-INFINITY, -1);
+  }
+  while (k) {
+    ThreadGetTopK<T, MaxLength, BlockSize>(topk, beam, k,
+                                           src + blockIdx.x * lds, firststep,
+                                           is_empty, max, dim, tid);
+
+    sh_topk[tid] = topk[0];
+    BlockReduce<T, MaxLength, BlockSize>(sh_topk, maxid, topk, &output,
+                                         &indices, beam, k, tid, warp);
+  }
+}
+
+template <typename T>
+class TopkOpCUDAKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
+                   "It must use GPUPlace.");
+    auto* input = ctx.Input<Tensor>("X");
+    auto* output = ctx.Output<Tensor>("Out");
+    auto* indices = ctx.Output<Tensor>("Indices");
+    size_t k = static_cast<int>(ctx.Attr<int>("k"));
+
+    const T* input_data = input->data<T>();
+
+    T* output_data = output->mutable_data<T>(ctx.GetPlace());
+    // FIXME(typhoonzero): data is always converted to type T?
+    int* indices_data = indices->mutable_data<int>(ctx.GetPlace());
+
+    size_t input_height = input->dims()[0];
+    size_t input_width = input->dims()[1];
+    if (k > input_width) k = input_width;
+
+    // NOTE: pass lds and dim same to input width.
+    // NOTE: old matrix implementation of stride is different to eigen.
+    // TODO(typhoonzero): launch kernel on specified stream.
+    // TODO(typhoonzero): refine this kernel.
+    dim3 threads(256, 1);
+    dim3 grid(input_height, 1);
+
+    KeMatrixTopK<T, 5, 256><<<grid, threads>>>(
+        output_data, output->dims()[1], indices_data, input_data, input_width,
+        input_width, int(k));
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OP_GPU_KERNEL(top_k, paddle::operators::TopkOpCUDAKernel<float>);

paddle/operators/top_k_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include <algorithm>
+#include <iostream>
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
+
+template <typename Place, typename T>
+class TopkKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    // Get the top k elements of each row of input tensor
+    // FIXME: only deal with matrix(2d tensor).
+    auto* input = ctx.Input<Tensor>("X");
+    auto* output = ctx.Output<Tensor>("Out");
+    auto* indices = ctx.Output<Tensor>("Indices");
+    // k is determined by Attr
+    const size_t k = static_cast<int>(ctx.Attr<int>("k"));
+
+    T* output_data = output->mutable_data<T>(ctx.GetPlace());
+    T* indices_data = indices->mutable_data<T>(ctx.GetPlace());
+
+    auto eg_input = EigenMatrix<T>::From(*input);
+
+    // reshape input to a flattern matrix(like flat_inner_dims)
+    framework::DDim inputdims = input->dims();
+    const size_t row = framework::product(
+        framework::slice_ddim(inputdims, 0, inputdims.size() - 1));
+    const size_t col = inputdims[inputdims.size() - 1];
+    Eigen::DSizes<int, 2> flat2dims(row, col);
+    // NOTE: eigen shape doesn't affect paddle tensor.
+    eg_input.reshape(flat2dims);
+
+    for (size_t i = 0; i < row; i++) {
+      std::vector<std::pair<T, size_t>> vec;
+      for (size_t j = 0; j < col; j++) {
+        vec.push_back(std::pair<T, size_t>(eg_input(i, j), j));
+      }
+
+      std::partial_sort(
+          vec.begin(), vec.begin() + k, vec.end(),
+          [](const std::pair<T, size_t>& l, const std::pair<T, size_t>& r) {
+            return l.first > r.first;
+          });
+      for (size_t j = 0; j < k; j++) {
+        output_data[i * k + j] = vec[j].first;
+        indices_data[i * k + j] = vec[j].second;
+      }
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/pybind/pybind.cc

@@ -49,6 +49,8 @@ USE_OP(minus);
 USE_OP(cos_sim);
 USE_CPU_ONLY_OP(gather);
 USE_CPU_ONLY_OP(scatter);
+USE_OP(top_k);
+USE_OP(squared_l2_distance);
 
 namespace paddle {
 namespace framework {

paddle/scripts/docker/build.sh

@@ -37,7 +37,7 @@ Configuring cmake in /paddle/build ...
       -DWITH_PYTHON=${WITH_PYTHON:-ON}
       -DWITH_SWIG_PY=${WITH_SWIG_PY:-ON}
       -DCUDNN_ROOT=/usr/
-      -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-OFF}
+      -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-ON}
       -DWITH_TESTING=${WITH_TESTING:-ON}
       -DCMAKE_EXPORT_COMPILE_COMMANDS=ON
 ========================================

paddle/scripts/docker/build_android.sh

@@ -2,22 +2,58 @@
 
 set -xe
 
-mkdir -p /paddle/build_android
-cd /paddle/build_android
-rm -rf /paddle/install 2>/dev/null || true
-cmake -DCMAKE_SYSTEM_NAME=Android \
-      -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_STANDALONE_TOOLCHAIN \
-      -DANDROID_ABI=armeabi-v7a \
-      -DANDROID_ARM_NEON=ON \
-      -DANDROID_ARM_MODE=ON \
-      -DHOST_C_COMPILER=/usr/bin/gcc \
-      -DHOST_CXX_COMPILER=/usr/bin/g++ \
-      -DCMAKE_INSTALL_PREFIX=/paddle/install \
-      -DCMAKE_BUILD_TYPE=RelWithDebInfo \
-      -DCMAKE_C_FLAGS_RELWITHDEBINFO="-O3" \
-      -DCMAKE_CXX_FLAGS_RELWITHDEBINFO="-O3" \
-      -DWITH_C_API=ON \
-      -DWITH_SWIG_PY=OFF \
-      ..
+BUILD_ROOT=/paddle/build_android
+DEST_ROOT=/paddle/install
+
+rm -rf $BUILD_ROOT 2>/dev/null || true
+mkdir -p $BUILD_ROOT
+cd $BUILD_ROOT
+
+if [ $ANDROID_ABI == "armeabi-v7a" ]; then
+  cmake -DCMAKE_SYSTEM_NAME=Android \
+        -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_ARM_STANDALONE_TOOLCHAIN \
+        -DANDROID_ABI=$ANDROID_ABI \
+        -DANDROID_ARM_NEON=ON \
+        -DANDROID_ARM_MODE=ON \
+        -DHOST_C_COMPILER=/usr/bin/gcc \
+        -DHOST_CXX_COMPILER=/usr/bin/g++ \
+        -DCMAKE_INSTALL_PREFIX=$DEST_ROOT \
+        -DCMAKE_BUILD_TYPE=Release \
+        -DUSE_EIGEN_FOR_BLAS=ON \
+        -DWITH_C_API=ON \
+        -DWITH_SWIG_PY=OFF \
+        -DWITH_STYLE_CHECK=OFF \
+        ..
+elif [ $ANDROID_ABI == "arm64-v8a" ]; then
+  cmake -DCMAKE_SYSTEM_NAME=Android \
+        -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_ARM64_STANDALONE_TOOLCHAIN \
+        -DANDROID_ABI=$ANDROID_ABI \
+        -DANDROID_ARM_MODE=ON \
+        -DHOST_C_COMPILER=/usr/bin/gcc \
+        -DHOST_CXX_COMPILER=/usr/bin/g++ \
+        -DCMAKE_INSTALL_PREFIX=$DEST_ROOT \
+        -DCMAKE_BUILD_TYPE=Release \
+        -DUSE_EIGEN_FOR_BLAS=OFF \
+        -DWITH_C_API=ON \
+        -DWITH_SWIG_PY=OFF \
+        -DWITH_STYLE_CHECK=OFF \
+        ..
+elif [ $ANDROID_ABI == "armeabi" ]; then
+  cmake -DCMAKE_SYSTEM_NAME=Android \
+        -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_ARM_STANDALONE_TOOLCHAIN \
+        -DANDROID_ABI=$ANDROID_ABI \
+        -DANDROID_ARM_MODE=ON \
+        -DHOST_C_COMPILER=/usr/bin/gcc \
+        -DHOST_CXX_COMPILER=/usr/bin/g++ \
+        -DCMAKE_INSTALL_PREFIX=/paddle/install \
+        -DCMAKE_BUILD_TYPE=Release \
+        -DWITH_C_API=ON \
+        -DWITH_SWIG_PY=OFF \
+        -DWITH_STYLE_CHECK=OFF \
+        ..
+else
+  echo "Invalid ANDROID_ABI: $ANDROID_ABI"
+fi
+
 make -j `nproc`
 make install -j `nproc`

paddle/scripts/travis/build_android.sh

@@ -22,6 +22,7 @@ cmake -DCMAKE_SYSTEM_NAME=Android \
       -DANDROID_ABI=armeabi-v7a \
       -DANDROID_ARM_NEON=ON \
       -DANDROID_ARM_MODE=ON \
+      -DUSE_EIGEN_FOR_BLAS=ON \
       -DWITH_C_API=ON \
       -DWITH_SWIG_PY=OFF \
       -DWITH_STYLE_CHECK=OFF \

proto/ModelConfig.proto

@@ -287,6 +287,11 @@ message PadConfig {
   repeated uint32 pad_w = 4;
 }
 
+message ReshapeConfig {
+  repeated uint32 height_axis = 1;
+  repeated uint32 width_axis = 2;
+}
+
 message MultiBoxLossConfig {
   required uint32 num_classes = 1;
   required float overlap_threshold = 2;
@@ -339,7 +344,6 @@ message LayerInputConfig {
 }
 
 message LayerConfig {
-
   required string name = 1;
   required string type = 2;
   optional uint64 size = 3;
@@ -516,6 +520,9 @@ message LayerConfig {
   optional double delta = 57 [ default = 1.0 ];
 
   optional uint64 depth = 58 [ default = 1 ];
+
+  // for switch order layer
+  optional ReshapeConfig reshape_conf = 59;
 }
 
 message EvaluatorConfig {

python/paddle/trainer/config_parser.py

@@ -3670,6 +3670,15 @@ class RecurrentLayerGroup(LayerBase):
             name, 'recurrent_layer_group', 0, inputs=[], device=device)
 
 
+@config_layer('switch_order')
+class SwitchOrderLayer(LayerBase):
+    def __init__(self, name, inputs, reshape, **xargs):
+        super(SwitchOrderLayer, self).__init__(
+            name, 'switch_order', 0, inputs=inputs, **xargs)
+        self.config.reshape_conf.heightAxis.extend(reshape['height'])
+        self.config.reshape_conf.widthAxis.extend(reshape['width'])
+
+
 # Deprecated, use a new layer specific class instead
 @config_func
 def Layer(name, type, **xargs):

python/paddle/trainer_config_helpers/layers.py

@@ -131,6 +131,7 @@ __all__ = [
     'row_conv_layer',
     'dropout_layer',
     'prelu_layer',
+    'switch_order_layer',
     'gated_unit_layer',
     'crop_layer',
     'sub_nested_seq_layer',
@@ -239,6 +240,7 @@ class LayerType(object):
     SMOOTH_L1 = 'smooth_l1'
 
     PRELU = 'prelu'
+    SWITCH_ORDER_LAYER = 'switch_order'
     CROP_LAYER = 'crop'
     SUB_NESTED_SEQ = 'sub_nested_seq'
     CLIP_LAYER = 'clip'
@@ -6404,6 +6406,48 @@ def gated_unit_layer(input,
         layer_attr=layer_attr)
 
 
+@layer_support()
+@wrap_name_default('switch_order')
+def switch_order_layer(input,
+                       name=None,
+                       reshape=None,
+                       act=None,
+                       layer_attr=None):
+    """
+    This layer switch dimension order of image input. 
+    From order "batchSize, channels, height, width"
+    to order "batchSize, height, width, channels".
+
+    The example usage is:
+
+    .. code-block:: python
+       reshape = {'height':[ 0, 1, 2], 'width':[3]}
+       switch = switch_order(input=layer, name='switch', reshape=reshape)
+
+    :param input: The input layer.
+    :type input: LayerOutput
+    :param name: Name of this layer.
+    :type name: basestring
+    :param reshape: reshape matrix by axises.
+    :type reshape: Dict
+    :return: LayerOutput object.
+    :rtype: LayerOutput
+    """
+    assert isinstance(input, LayerOutput)
+    l = Layer(
+        name=name,
+        inputs=input.name,
+        reshape=reshape,
+        type=LayerType.SWITCH_ORDER_LAYER,
+        active_type=act.name,
+        **ExtraLayerAttribute.to_kwargs(layer_attr))
+    return LayerOutput(
+        name=name,
+        layer_type=LayerType.SWITCH_ORDER_LAYER,
+        parents=input,
+        size=l.config.size)
+
+
 @wrap_name_default()
 @layer_support()
 def crop_layer(input, offset, axis=2, shape=None, name=None, layer_attr=None):

python/paddle/v2/framework/op.py

@@ -4,8 +4,8 @@ import paddle.v2.framework.proto.framework_pb2 as framework_pb2
 
 def get_all_op_protos():
     """
-    Get all registered op proto from Paddle C++
-    :return: list of OpProto
+    Get all registered op proto from PaddlePaddle C++ end.
+    :return: A list of registered OpProto.
     """
     protostrs = core.get_all_op_protos()
     ret_values = []
@@ -21,8 +21,8 @@ def is_str(s):
 
 class OpDescCreationMethod(object):
     """
-    A Functor object to convert user input(use key word args) to OpDesc based on
-    OpProto.
+    Convert the user's input(only keyword arguments are supported) to OpDesc
+    based on the OpProto.
 
     :param op_proto: The OpProto object.
     :type op_proto: op_proto_pb2.OpProto
@@ -30,17 +30,18 @@ class OpDescCreationMethod(object):
 
     def __init__(self, op_proto):
         if not isinstance(op_proto, framework_pb2.OpProto):
-            raise TypeError("Argument should be OpProto")
+            raise TypeError(
+                "Type of op_proto should be OpProto in PaddlePaddle.")
         self.__op_proto__ = op_proto
 
     def __call__(self, *args, **kwargs):
         """
-        Convert user input to OpDesc. Only key-word args are supported. 
-        :return: OpDesc based on user input
+        Convert user's input to OpDesc. Only keyword arguments are supported.
+        :return: The OpDesc based on user input.
         :rtype: op_desc_pb2.OpDesc
         """
         if len(args) != 0:
-            raise ValueError("Only keyword arguments is supported by Paddle")
+            raise ValueError("Only keyword arguments are supported.")
         op_desc = framework_pb2.OpDesc()
 
         for input_parameter in self.__op_proto__.inputs:
@@ -49,8 +50,9 @@ class OpDescCreationMethod(object):
                 input_arguments = [input_arguments]
 
             if not input_parameter.duplicable and len(input_arguments) > 1:
-                raise ValueError("Input %s only accepts one input, but give %d"
-                                 % (input_parameter.name, len(input_arguments)))
+                raise ValueError(
+                    "Input %s expects only one input, but %d are given." %
+                    (input_parameter.name, len(input_arguments)))
 
             ipt = op_desc.inputs.add()
             ipt.parameter = input_parameter.name
@@ -63,7 +65,7 @@ class OpDescCreationMethod(object):
 
             if not output_parameter.duplicable and len(output_arguments) > 1:
                 raise ValueError(
-                    "Output %s only accepts one output, but give %d" %
+                    "Output %s expects only one output, but %d are given." %
                     (output_parameter.name, len(output_arguments)))
 
             out = op_desc.outputs.add()
@@ -100,15 +102,17 @@ class OpDescCreationMethod(object):
                         pair.first = p[0]
                         pair.second = p[1]
                 else:
-                    raise NotImplementedError("Not support attribute type " +
-                                              str(attr.type))
+                    raise NotImplementedError(
+                        "A not supported attribute type: %s." % (
+                            str(attr.type)))
 
         return op_desc
 
     @staticmethod
     def any_is_true(generator):
         """
-        Reduce a bool array to one. If any of them is True, then return True.
+        Reduce a boolean array to a single boolean parameter. If any element in
+        the array is True, this function will return True, otherwise False.
         """
         for flag in generator:
             if flag:
@@ -127,7 +131,7 @@ class OpInfo(object):
 
 def create_op_creation_method(op_proto):
     """
-    Generate op creation method for an OpProto
+    Generate op creation method for an OpProto.
     """
     method = OpDescCreationMethod(op_proto)
 
@@ -146,20 +150,23 @@ def create_op_creation_method(op_proto):
 class OperatorFactory(object):
     def __init__(self):
         self.op_methods = dict()
+
         for op_proto in get_all_op_protos():
             method = create_op_creation_method(op_proto)
             self.op_methods[method.name] = method
 
     def __call__(self, *args, **kwargs):
-        if 'type' in kwargs:
+        if "type" in kwargs:
             if len(args) != 0:
-                raise ValueError("All Paddle argument should be key-word "
-                                 "argument except type")
-            t = kwargs.pop('type')
+                raise ValueError(
+                    "Except the argument \"type\","
+                    "all of the other arguments should be keyword arguments.")
+            t = kwargs.pop("type")
         else:
             if len(args) != 1:
-                raise ValueError("All Paddle argument should be key-word "
-                                 "argument except type")
+                raise ValueError(
+                    "Except the argument \"type\","
+                    "all of the other arguments should be keyword arguments.")
             t = args[0]
 
         return self.get_op_info(t).method(**kwargs)
@@ -169,7 +176,7 @@ class OperatorFactory(object):
 
     def get_op_info(self, t):
         if t not in self.op_methods:
-            raise ValueError("operator %s is not registered", t)
+            raise ValueError("The operator: %s is not registered." % t)
         return self.op_methods.get(t)
 
     def get_op_input_names(self, type):
@@ -184,7 +191,7 @@ class OperatorFactory(object):
 
 class __RecurrentOp__(object):
     __proto__ = None
-    type = 'recurrent'
+    type = "recurrent"
 
     def __init__(self):
         # cache recurrent_op's proto
@@ -194,8 +201,8 @@ class __RecurrentOp__(object):
                     self.__proto__ = op_proto
 
     def __call__(self, *args, **kwargs):
-        if self.type not in args and 'type' not in kwargs:
-            kwargs['type'] = self.type
+        if self.type not in args and "type" not in kwargs:
+            kwargs["type"] = self.type
         # create proto
         create_method = OpDescCreationMethod(self.__proto__)
         proto = create_method(*args, **kwargs)
@@ -203,5 +210,5 @@ class __RecurrentOp__(object):
         return core.RecurrentOp.create(proto.SerializeToString())
 
 
-Operator = OperatorFactory()  # Default global factory
+Operator = OperatorFactory()  # The default global factory
 RecurrentOp = __RecurrentOp__()

python/paddle/v2/framework/tests/CMakeLists.txt

@@ -17,6 +17,7 @@ py_test(test_cross_entropy_op SRCS test_cross_entropy_op.py)
 py_test(test_gather_op SRCS test_gather_op.py)
 py_test(test_scatter_op SRCS test_scatter_op.py)
 py_test(test_fill_zeros_like_op SRCS test_fill_zeros_like_op.py)
+py_test(test_top_k_op SRCS test_top_k_op.py)
 
 py_test(gradient_checker SRCS gradient_checker.py)
 
@@ -33,3 +34,4 @@ py_test(test_gradient_checker SRCS test_gradient_checker.py)
 py_test(test_lookup_table SRCS test_lookup_table.py)
 py_test(test_scale_and_identity_op SRCS test_scale_and_identity_op.py)
 py_test(mnist SRCS mnist.py)
+py_test(test_squared_l2_distance_op SRCS test_squared_l2_distance_op.py)

python/paddle/v2/framework/tests/mnist.py

@@ -38,9 +38,9 @@ def feed_data(name, data):
     assert isinstance(data, numpy.ndarray)
     tensor = scope.find_var(name).get_tensor()
     tensor.set_dims(data.shape)
-    if data.dtype == numpy.dtype('int32'):
+    if data.dtype == numpy.dtype("int32"):
         tensor.alloc_int(place)
-    elif data.dtype == numpy.dtype('float32'):
+    elif data.dtype == numpy.dtype("float32"):
         tensor.alloc_float(place)
     else:
         raise ValueError("data type not supported")
@@ -74,22 +74,25 @@ def init_param(net, param_name, dims):
 # fc_layer
 def fc_layer(net, input, size, act="softmax", bias=True, param=None, name=None):
     """
-    Add a fc layer to net
+    The fully connected layer.
 
-    :param input: input variable name.
+    :param input: The name of input variable.
     :type input: str
-    :param size: fully connected layer size.
-    :param act: activation name
-    :param param: parameter attribute, used for initialize parameters.
-    :param bias: bias attribute. False will not have a bias.
-    :param name: the name of fc layer. If not set, model will generate a
-    readable name
-    :return: output variable name.
+    :param size: The size of fully connected layer.
+    :param act: The name of activation.
+    :param param: The attribute of learnable parameter which can be used to
+                  modify initialization mean and std of the parameter.
+    :param bias: The attribute of bias. If set False, this layer does not have
+                 a bias.
+    :param name: The name of this layer. If it is not set explictly, a name
+                 will be generated automatically.
+    :return: The name of the output variable.
     """
+
     if name is None:
-        name = 'fc_%d' % uniq_id()
+        name = "fc_%d" % uniq_id()
     if not isinstance(name, str):
-        raise ValueError("name should be string")
+        raise ValueError("The name of a layer should be a string.")
 
     input_dims = scope.find_var(input).get_tensor().get_dims()
 
@@ -123,7 +126,7 @@ def fc_layer(net, input, size, act="softmax", bias=True, param=None, name=None):
 
 
 def cross_entropy_layer(net, input, label):
-    cost_name = 'cross_entropy_%d' % uniq_id()
+    cost_name = "cross_entropy_%d" % uniq_id()
     cross_entropy_op = Operator(
         "onehot_cross_entropy", X=input, label=label, Y=cost_name)
     net.append_op(cross_entropy_op)
@@ -177,8 +180,8 @@ def error_rate(predict, label):
     return error_num / float(len(label))
 
 
-images = data_layer(name='pixel', dims=[BATCH_SIZE, 784])
-labels = data_layer(name='label', dims=[BATCH_SIZE])
+images = data_layer(name="pixel", dims=[BATCH_SIZE, 784])
+labels = data_layer(name="label", dims=[BATCH_SIZE])
 fc1 = fc_layer(net=forward_net, input=images, size=100, act="sigmoid")
 fc2 = fc_layer(net=forward_net, input=fc1, size=100, act="sigmoid")
 predict = fc_layer(net=forward_net, input=fc2, size=10, act="softmax")

python/paddle/v2/framework/tests/op_test_util.py

@@ -66,7 +66,7 @@ class OpTestMeta(type):
                     self.assertTrue(
                         numpy.allclose(
                             actual, expect, atol=1e-05),
-                        "output name: " + out_name + "has diff")
+                        "output name: " + out_name + " has diff")
 
         obj.test_all = test_all
         return obj