Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into fix-3789

doc/api/v2/config/layer.rst

@@ -434,9 +434,9 @@ lambda_cost
 ..  autoclass:: paddle.v2.layer.lambda_cost
     :noindex:
 
-mse_cost
+square_error_cost
 --------
-..  autoclass:: paddle.v2.layer.mse_cost
+..  autoclass:: paddle.v2.layer.square_error_cost
     :noindex:
 
 rank_cost

doc/design/graph.md

@@ -0,0 +1,70 @@
+# Design Doc: Computations as a Graph
+
+A primary goal of the refactorization of PaddlePaddle is a more flexible representation of deep learning computation, in particular, a graph of operators and variables, instead of sequences of layers as before.
+
+This document explains that the construction of a graph as three steps:
+
+- construct the forward part
+- construct the backward part
+- construct the optimization part
+
+## The Construction of a Graph
+
+Let us take the problem of image classification as a simple example.  The application program that trains the model looks like:
+
+```python
+x = layer.data("images")
+l = layer.data("label")
+y = layer.fc(x)
+cost = layer.mse(y, l)
+optimize(cost)
+train(cost, reader=mnist.train())
+```
+
+### Forward Part
+
+The first four lines of above program build the forward part of the graph.
+
+![](images/graph_construction_example_forward_only.png)
+
+In particular, the first line `x = layer.data("images")` creates variable x and a Feed operator that copies a column from the minibatch to x.  `y = layer.fc(x)` creates not only the FC operator and output variable y, but also two parameters, W and b, and the initialization operators.
+
+Initialization operators are kind of "run-once" operators -- the `Run` method increments a class data member counter so to run at most once.  By doing so, a parameter wouldn't be initialized repeatedly, say, in every minibatch.
+
+In this example, all operators are created as `OpDesc` protobuf messages, and all variables are `VarDesc`.  These protobuf messages are saved in a `BlockDesc` protobuf message.
+
+### Backward Part
+
+The fifth line `optimize(cost)` calls two functions, `ConstructBackwardGraph` and `ConstructOptimizationGraph`.
+
+`ConstructBackwardGraph` traverses the forward graph in the `BlockDesc` protobuf message and builds the backward part.
+
+![](images/graph_construction_example_forward_backward.png)
+
+According to the chain rule of gradient computation, `ConstructBackwardGraph` would
+
+1. create a gradient operator G for each operator F,
+1. make all inputs, outputs, and outputs' gradient of F as inputs of G,
+1. create gradients for all inputs of F, except for those who don't have gradients, like x and l, and
+1. make all these gradients as outputs of G.
+
+### Optimization Part
+
+For each parameter, like W and b created by `layer.fc`, marked as double circles in above graphs, `ConstructOptimizationGraph` creates an optimization operator to apply its gradient.  Here results in the complete graph:
+
+![](images/graph_construction_example_all.png)
+
+## Block and Graph
+
+The word block and graph are interchangable in the desgin of PaddlePaddle.  A [Block[(https://github.com/PaddlePaddle/Paddle/pull/3708) is a metaphore of the code and local variables in a pair of curly braces in programming languages, where operators are like statements or instructions.  A graph of operators and variables is a representation of the block.
+
+A Block keeps operators in an array `BlockDesc::ops`
+
+```protobuf
+message BlockDesc {
+  repeated OpDesc ops = 1;
+  repeated VarDesc vars = 2;
+}
+```
+
+in the order that there appear in user programs, like the Python program at the beginning of this article.  We can imagine that in `ops`,  we have some forward operators, followed by some gradient operators, and then some optimization operators.

doc/design/images/graph_construction_example.bash

@@ -0,0 +1,11 @@
+cat ./graph_construction_example.dot | \
+    sed 's/color=red/color=red, style=invis/g' | \
+    sed 's/color=green/color=green, style=invis/g' | \
+    dot -Tpng > graph_construction_example_forward_only.png
+
+cat ./graph_construction_example.dot | \
+    sed 's/color=green/color=green, style=invis/g' | \
+    dot -Tpng > graph_construction_example_forward_backward.png
+
+cat ./graph_construction_example.dot | \
+    dot -Tpng > graph_construction_example_all.png

doc/design/images/graph_construction_example.dot

@@ -0,0 +1,69 @@
+digraph ImageClassificationGraph {
+        ///////// The forward part /////////
+        FeedX [label="Feed", color=blue, shape=box];
+        FeedY [label="Feed", color=blue, shape=box];
+        InitW [label="Init", color=blue, shape=diamond];
+        Initb [label="Init", color=blue, shape=diamond];
+        FC [label="FC", color=blue, shape=box];
+        MSE [label="MSE", color=blue, shape=box];
+
+        x [label="x", color=blue, shape=oval];
+        l [label="l", color=blue, shape=oval];
+        y [label="y", color=blue, shape=oval];
+        W [label="W", color=blue, shape=doublecircle];
+        b [label="b", color=blue, shape=doublecircle];
+        cost [label="cost", color=blue, shape=oval];
+
+        FeedX -> x -> FC -> y -> MSE -> cost [color=blue];
+        FeedY -> l [color=blue];
+        InitW -> W [color=blue];
+        Initb -> b [color=blue];
+        W -> FC [color=blue];
+        b -> FC [color=blue];
+        l -> MSE [color=blue];
+
+        ////////// The backward part /////////
+        MSE_Grad [label="MSE_grad", color=red, shape=box];
+        FC_Grad [label="FC_grad", color=red, shape=box];
+
+        d_cost [label="d cost", color=red, shape=oval];
+        d_y [label="d y", color=red, shape=oval];
+        d_b [label="d b", color=red, shape=oval];
+        d_W [label="d W", color=red, shape=oval];
+
+        cost -> MSE_Grad [color=red];
+        d_cost -> MSE_Grad [color=red];
+        x -> MSE_Grad [color=red];
+        l -> MSE_Grad [color=red];
+        y -> MSE_Grad -> d_y [color=red];
+
+        x -> FC_Grad [color=red];
+        y -> FC_Grad [color=red];
+        d_y -> FC_Grad [color=red];
+        W -> FC_Grad -> d_W [color=red];
+        b -> FC_Grad -> d_b [color=red];
+
+        ////////// The optimizaiton part //////////
+
+        OPT_W [label="SGD", color=green, shape=box];
+        OPT_b [label="SGD", color=green, shape=box];
+
+        W -> OPT_W [color=green];
+        b -> OPT_b [color=green];
+        d_W -> OPT_W -> W [color=green];
+        d_b -> OPT_b -> b [color=green];
+
+        ////////// Groupings //////////
+
+        subgraph clusterMSE {
+                style=invis;
+                MSE;
+                MSE_Grad;
+        }
+
+        subgraph clusterFC {
+                style=invis;
+                FC;
+                FC_Grad;
+        }
+}

doc/design/simple_op_design.md

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

doc/getstarted/basic_usage/index_cn.rst

@@ -55,7 +55,7 @@ PaddlePaddle是源于百度的一个深度学习平台。这份简短的介绍
     # 线性计算网络层: ȳ = wx + b
     ȳ = fc_layer(input=x, param_attr=ParamAttr(name='w'), size=1, act=LinearActivation(), bias_attr=ParamAttr(name='b'))
     # 计算误差函数，即  ȳ 和真实 y 之间的距离
-    cost = mse_cost(input= ȳ, label=y)
+    cost = square_error_cost(input= ȳ, label=y)
     outputs(cost)
 
 
@@ -69,7 +69,7 @@ PaddlePaddle是源于百度的一个深度学习平台。这份简短的介绍
     
     - **数据层**：数据层 `data_layer` 是神经网络的入口，它读入数据并将它们传输到接下来的网络层。这里数据层有两个，分别对应于变量 `x` 和 `y`。
     - **全连接层**：全连接层 `fc_layer` 是基础的计算单元，这里利用它建模变量之间的线性关系。计算单元是神经网络的核心，PaddlePaddle支持大量的计算单元和任意深度的网络连接，从而可以拟合任意的函数来学习复杂的数据关系。
-    - **回归误差代价层**：回归误差代价层 `mse_cost` 是众多误差代价函数层的一种，它们在训练过程作为网络的出口，用来计算模型的误差，是模型参数优化的目标函数。
+    - **回归误差代价层**：回归误差代价层 `square_error_cost` 是众多误差代价函数层的一种，它们在训练过程作为网络的出口，用来计算模型的误差，是模型参数优化的目标函数。
 
 定义了网络结构并保存为 `trainer_config.py` 之后，运行以下训练命令：
 

doc/getstarted/basic_usage/index_en.rst

@@ -49,7 +49,7 @@ To recover this relationship between ``X`` and ``Y``, we use a neural network wi
         x = data_layer(name='x', size=1)
         y = data_layer(name='y', size=1)
         y_predict = fc_layer(input=x, param_attr=ParamAttr(name='w'), size=1, act=LinearActivation(), bias_attr=ParamAttr(name='b'))
-        cost = mse_cost(input=y_predict, label=y)
+        cost = square_error_cost(input=y_predict, label=y)
         outputs(cost)
 
 Some of the most fundamental usages of PaddlePaddle are demonstrated:

doc/getstarted/concepts/src/train.py

@@ -8,7 +8,7 @@ paddle.init(use_gpu=False)
 x = paddle.layer.data(name='x', type=paddle.data_type.dense_vector(2))
 y_predict = paddle.layer.fc(input=x, size=1, act=paddle.activation.Linear())
 y = paddle.layer.data(name='y', type=paddle.data_type.dense_vector(1))
-cost = paddle.layer.mse_cost(input=y_predict, label=y)
+cost = paddle.layer.square_error_cost(input=y_predict, label=y)
 
 # create parameters
 parameters = paddle.parameters.create(cost)

doc/getstarted/concepts/use_concepts_cn.rst

@@ -81,9 +81,9 @@ PaddlePaddle支持不同类型的输入数据，主要包括四种类型，和
 ..	code-block:: bash
 
     y_predict = paddle.layer.fc(input=x, size=1, act=paddle.activation.Linear())
-    cost = paddle.layer.mse_cost(input=y_predict, label=y)
+    cost = paddle.layer.square_error_cost(input=y_predict, label=y)
 
-其中，x与y为之前描述的输入层；而y_predict是接收x作为输入，接上一个全连接层；cost接收y_predict与y作为输入，接上均方误差层。
+其中，x与y为之前描述的输入层；而y_predict是接收x作为输入，接上一个全连接层；cost接收y_predict与y作为输入，接上平方误差层。
 
 最后一层cost中记录了神经网络的所有拓扑结构，通过组合不同的layer，我们即可完成神经网络的搭建。
 
@@ -147,4 +147,4 @@ PaddlePaddle支持不同类型的输入数据，主要包括四种类型，和
 ..  literalinclude:: src/train.py
     :linenos:
 
-有关线性回归的实际应用，可以参考PaddlePaddle book的 `第一章节 <http://book.paddlepaddle.org/index.html>`_。
+有关线性回归的实际应用，可以参考PaddlePaddle book的 `第一章节 <http://book.paddlepaddle.org/index.html>`_。

doc/howto/dev/use_eigen_cn.md

@@ -0,0 +1,146 @@
+## 在Paddle中如何使用Eigen
+
+神经网络本质上是一个计算图，计算需要的数据存放在`Tensor`中，而计算过程是由`Operartor`来描述的。在执行时，`Operator`调用对应`OpKernel`中的`Compute`接口，实现对`Tensor`的操作。
+
+
+### Eigen Tensor模块
+
+Eigen Tensor模块对element-wise计算提供了强大的支持，并且书写一份代码，可以同时在CPU、GPU执行。但Eigen Tensor是一个正在开发中的模块，因此可能测试不够完备，文档较少。
+
+关于Eigen Tensor模块的详细介绍请参考[文档1](https://github.com/RLovelett/eigen/blob/master/unsupported/Eigen/CXX11/src/Tensor/README.md) 和[文档2](https://bitbucket.org/eigen/eigen/src/default/unsupported/Eigen/CXX11/src/Tensor/README.md)
+
+
+### paddle::framework::Tensor
+
+Paddle Tensor定义在framework目录下，其主要接口如下：
+
+```cpp
+class Tensor {
+ public:
+  /*! Return a pointer to mutable memory block. */
+  template <typename T>
+  inline T* data();
+  
+  /**
+   * @brief   Return a pointer to mutable memory block.
+   * @note    If not exist, then allocation.
+   */
+  template <typename T>
+  inline T* mutable_data(platform::Place place);
+  
+  /**
+   * @brief     Return a pointer to mutable memory block.
+   *
+   * @param[in] dims    The dimensions of the memory block.
+   * @param[in] place   The place of the memory block.
+   *
+   * @note      If not exist, then allocation.
+   */
+  template <typename T>
+  inline T* mutable_data(DDim dims, platform::Place place);
+  
+  /*! Resize the dimensions of the memory block. */
+  inline Tensor& Resize(const DDim& dims);
+  
+  /*! Return the dimensions of the memory block. */
+  inline const DDim& dims() const;
+
+ private:  
+  /*! holds the memory block if allocated. */
+  std::shared_ptr<Placeholder> holder_;
+  
+  /*! points to dimensions of memory block. */
+  DDim dim_;
+};
+```
+
+`Placeholder`的作用是延迟分配内存，即我们可以先定义一个Tensor，然后使用Resize接口设置Tensor的大小，最后再调用mutable_data接口分配实际的内存。
+
+```cpp
+paddle::framework::Tensor t;
+paddle::platform::CPUPlace place;
+// set size first
+t.Resize({2, 3});
+// allocate memory on CPU later
+t.mutable_data(place);
+```
+
+### paddle::framework::Tensor使用样例
+下面以AddOp为例说明Tensor的使用过程：
+
+- InferShape
+
+在运行神经网络计算图时，我们先调用每个`Operator`的`InferShape`接口，根据输入Tensor的大小来设置输出Tensor的大小，`Resize`接口会被调用。
+
+```cpp
+void InferShape(const framework::InferShapeContext &ctx) const override {
+  PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("X")->dims(),
+                    ctx.Input<Tensor>("Y")->dims(),
+                    "Two input of Add Op's dimension must be same.");
+  ctx.Output<Tensor>("Out")->Resize(ctx.Input<Tensor>("X")->dims());
+}
+```
+
+
+- Run
+
+`Operator`的`Run`接口最终会调用对应`OpKernel`的`Compute`接口，在这时真正的分配内存，`mutable_data`接口会被调用。
+
+```cpp
+void Compute(const framework::ExecutionContext& context) const override {
+  auto* input0 = context.Input<Tensor>("X");
+  auto* input1 = context.Input<Tensor>("Y");
+  auto* output = context.Output<Tensor>("Out");
+
+  output->mutable_data<T>(context.GetPlace());
+
+  auto x = EigenVector<T>::Flatten(*input0);
+  auto y = EigenVector<T>::Flatten(*input1);
+  auto z = EigenVector<T>::Flatten(*output);
+
+  auto place = context.GetEigenDevice<Place>();
+
+  z.device(place) = x + y;
+}
+```
+
+
+### paddle::framework::Tensor到EigenTensor的转换
+
+如上一小节所示，在具体的计算中，我们需要先把输入Tensor和输出Tensor转换为Eigen支持的格式。我们在[eigen.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/eigen.h)中提供了一些全局函数用来实现paddle::framework::Tensor到EigenTensor/EigenMatrix/EigenVector/EigenScalar的转换。
+
+以EigenTensor为例，做一个介绍
+
+```cpp
+Tensor t;
+float* p = t.mutable_data<float>(make_ddim({1, 2, 3}), platform::CPUPlace());
+for (int i = 0; i < 1 * 2 * 3; i++) {
+  p[i] = static_cast<float>(i);
+}
+
+EigenTensor<float, 3>::Type et = EigenTensor<float, 3>::From(t);
+```
+
+From是EigenTensor模板提供的一个接口，可以实现从paddle::framework::Tensor到对EigenTensor的转换。由于Tensor的rank是模板参数，因此在转换时需要显示的指定。
+
+在Eigen中，不同rank的Tensor是不同类型，Vector是rank为1的Tensor。需要额外注意的是，EigenVector<T>::From方法是把paddle中的一维Tensor转为Eigen的一维Tensor，在这里用EigenVector来表示；而EigenVector<T>::Flatten方法是把paddle中的一个Tensor进行reshape操作，压扁成为Eigen的一维Tensor，类型仍然为EigenVector。
+
+更多的转换方法请参考eigen_test.cc中的[单元测试](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/eigen_test.cc)。
+
+
+
+### 实现计算
+
+当需要完成计算时，我们需要等式左边的EigenTensor调用device接口。在这里需要注意的是，这里的EigenTensor之间的运算只是改变了原有Tensor中的数据，而不会改变原有Tensor的shape信息。
+
+```cpp
+auto x = EigenVector<T>::Flatten(*input0);
+auto y = EigenVector<T>::Flatten(*input1);
+auto z = EigenVector<T>::Flatten(*output);
+auto place = context.GetEigenDevice<Place>();
+z.device(place) = x + y;
+```
+
+在这段代码中，input0/input1/output可以是任意维度的Tensor。我们调用了EigenVector的Flatten接口，把任意维度的Tensor转为了一维的EigenVector。而在计算结束之后，input0/input1/output的原有shape信息不变。如果想改变原有Tensor的shape信息，可以调用Resize接口进行改变。
+
+由于Eigen Tensor模块的文档较少，我们可以参考TensorFlow的[kernels](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/core/kernels)模块下的相关`OpKernel`的计算代码。

doc/howto/usage/k8s/k8s_distributed_cn.md

@@ -213,7 +213,7 @@ I1116 09:10:17.123440    50 Util.cpp:130] Calling runInitFunctions
 I1116 09:10:17.123764    50 Util.cpp:143] Call runInitFunctions done.
 [WARNING 2016-11-16 09:10:17,227 default_decorators.py:40] please use keyword arguments in paddle config.
 [INFO 2016-11-16 09:10:17,239 networks.py:1282] The input order is [movie_id, title, genres, user_id, gender, age, occupation, rating]
-[INFO 2016-11-16 09:10:17,239 networks.py:1289] The output order is [__mse_cost_0__]
+[INFO 2016-11-16 09:10:17,239 networks.py:1289] The output order is [__square_error_cost_0__]
 I1116 09:10:17.392917    50 Trainer.cpp:170] trainer mode: Normal
 I1116 09:10:17.613910    50 PyDataProvider2.cpp:257] loading dataprovider dataprovider::process
 I1116 09:10:17.680917    50 PyDataProvider2.cpp:257] loading dataprovider dataprovider::process

paddle/framework/attribute.cc

@@ -43,6 +43,10 @@ template <>
 AttrType AttrTypeID<std::vector<std::string>>() {
   return STRINGS;
 }
+template <>
+AttrType AttrTypeID<std::vector<std::pair<int, int>>>() {
+  return INT_PAIRS;
+}
 
 Attribute GetAttrValue(const OpDesc::Attr& attr_desc) {
   switch (attr_desc.type()) {
@@ -76,6 +80,14 @@ Attribute GetAttrValue(const OpDesc::Attr& attr_desc) {
       }
       return val;
     }
+    case paddle::framework::AttrType::INT_PAIRS: {
+      std::vector<std::pair<int, int>> val(attr_desc.int_pairs_size());
+      for (int i = 0; i < attr_desc.int_pairs_size(); ++i) {
+        val[i].first = attr_desc.int_pairs(i).first();
+        val[i].second = attr_desc.int_pairs(i).second();
+      }
+      return val;
+    }
   }
   PADDLE_ENFORCE(false, "Unknown OpDesc::AttrDesc::type !");
   return boost::blank();

paddle/framework/attribute.h

@@ -28,7 +28,8 @@ namespace paddle {
 namespace framework {
 
 typedef boost::variant<boost::blank, int, float, std::string, std::vector<int>,
-                       std::vector<float>, std::vector<std::string>>
+                       std::vector<float>, std::vector<std::string>,
+                       std::vector<std::pair<int, int>>>
     Attribute;
 
 typedef std::unordered_map<std::string, Attribute> AttributeMap;
@@ -40,9 +41,9 @@ Attribute GetAttrValue(const OpDesc::Attr& attr_desc);
 
 // check whether a value(attribute) fit a certain limit
 template <typename T>
-class LargerThanChecker {
+class GreaterThanChecker {
  public:
-  explicit LargerThanChecker(T lower_bound) : lower_bound_(lower_bound) {}
+  explicit GreaterThanChecker(T lower_bound) : lower_bound_(lower_bound) {}
   void operator()(T& value) const {
     PADDLE_ENFORCE(value > lower_bound_, "larger_than check fail");
   }
@@ -109,8 +110,8 @@ class TypedAttrChecker {
     return *this;
   }
 
-  TypedAttrChecker& LargerThan(const T& lower_bound) {
-    value_checkers_.push_back(LargerThanChecker<T>(lower_bound));
+  TypedAttrChecker& GreaterThan(const T& lower_bound) {
+    value_checkers_.push_back(GreaterThanChecker<T>(lower_bound));
     return *this;
   }
 

paddle/framework/backward.cc

@@ -182,7 +182,7 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
                    });
 
     // process recurrent gradient op as a special operator.
-    if (forwardOp.Type() == "recurrent_op") {
+    if (forwardOp.Type() == "recurrent") {
       // NOTE clean up cycle call somewhere (RNN's stepnet constains itself), or
       // this will result in infinite loop.
       const auto& rnnop =

paddle/framework/backward.md

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

paddle/framework/backward_test.cc

@@ -148,14 +148,16 @@ class AddOpMaker : public OpProtoAndCheckerMaker {
 namespace f = paddle::framework;
 namespace ops = paddle::operators;
 using EnforceNotMet = paddle::platform::EnforceNotMet;
-REGISTER_OP(rowwise_add, f::NOP, f::RowWiseAddOpMaker, f::NOP);
-REGISTER_OP(mul, f::NOP, f::MulOpMaker, f::NOP);
-REGISTER_OP(sigmoid, f::NOP, f::SigmoidOpMaker, f::NOP);
+REGISTER_OP(rowwise_add, f::NOP, f::RowWiseAddOpMaker, rowwise_add_grad,
+            f::NOP);
+REGISTER_OP(mul, f::NOP, f::MulOpMaker, mul_grad, f::NOP);
+REGISTER_OP(sigmoid, f::NOP, f::SigmoidOpMaker, sigmoid_grad, f::NOP);
 REGISTER_OP_WITHOUT_GRADIENT(nograd, f::NOP, f::NoGradOpMaker);
 REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, f::NOP, f::FillZeroOpMaker);
-REGISTER_OP(add, f::NOP, f::AddOpMaker, f::NOP);
+REGISTER_OP(add, f::NOP, f::AddOpMaker, add_grad, f::NOP);
 REGISTER_OP_WITHOUT_GRADIENT(fc, f::FcOp, f::FcOpMaker);
-REGISTER_OP(many_output_op, f::NOP, f::ManyOutputOpMaker, f::NOP);
+REGISTER_OP(many_output_op, f::NOP, f::ManyOutputOpMaker, many_output_op_grad,
+            f::NOP);
 
 TEST(Backward, simple_op_grad) {
   auto fwd = f::OpRegistry::CreateOp(

paddle/framework/ddim.cc

@@ -21,16 +21,16 @@ namespace framework {
 /// @cond HIDDEN
 
 template <int i>
-Dim<i> make_dim(const int* d) {
+Dim<i> make_dim(const int64_t* d) {
   return Dim<i>(*d, make_dim<i - 1>(d + 1));
 }
 
 template <>
-Dim<1> make_dim<1>(const int* d) {
+Dim<1> make_dim<1>(const int64_t* d) {
   return Dim<1>(*d);
 }
 
-void make_ddim(DDim& ddim, const int* dims, int n) {
+void make_ddim(DDim& ddim, const int64_t* dims, int n) {
   switch (n) {
     case 1:
       ddim = make_dim<1>(dims);
@@ -67,13 +67,13 @@ void make_ddim(DDim& ddim, const int* dims, int n) {
 
 /// @endcond
 
-DDim make_ddim(std::initializer_list<int> dims) {
+DDim make_ddim(std::initializer_list<int64_t> dims) {
   DDim result(make_dim(0));
   make_ddim(result, dims.begin(), dims.size());
   return result;
 }
 
-DDim make_ddim(const std::vector<int>& dims) {
+DDim make_ddim(const std::vector<int64_t>& dims) {
   DDim result(make_dim(0));
   make_ddim(result, &dims[0], dims.size());
   return result;
@@ -81,12 +81,12 @@ DDim make_ddim(const std::vector<int>& dims) {
 
 /// @cond HIDDEN
 // XXX For some reason, putting this in an anonymous namespace causes errors
-class DynamicMutableIndexer : public boost::static_visitor<int&> {
+class DynamicMutableIndexer : public boost::static_visitor<int64_t&> {
  public:
   explicit DynamicMutableIndexer(int idx) : idx_(idx) {}
 
   template <int D>
-  int& operator()(Dim<D>& dim) const {
+  int64_t& operator()(Dim<D>& dim) const {
     return dim[idx_];
   }
 
@@ -94,12 +94,12 @@ class DynamicMutableIndexer : public boost::static_visitor<int&> {
   int idx_;
 };
 
-class DynamicConstIndexer : public boost::static_visitor<int> {
+class DynamicConstIndexer : public boost::static_visitor<int64_t> {
  public:
   explicit DynamicConstIndexer(int idx) : idx_(idx) {}
 
   template <int D>
-  int operator()(const Dim<D>& dim) const {
+  int64_t operator()(const Dim<D>& dim) const {
     return dim[idx_];
   }
 
@@ -109,22 +109,22 @@ class DynamicConstIndexer : public boost::static_visitor<int> {
 
 /// @endcond
 
-int& DDim::operator[](int idx) {
+int64_t& DDim::operator[](int idx) {
   return boost::apply_visitor(DynamicMutableIndexer(idx), var);
 }
 
-int DDim::operator[](int idx) const {
+int64_t DDim::operator[](int idx) const {
   return boost::apply_visitor(DynamicConstIndexer(idx), var);
 }
 
-ssize_t DDim::size() const { return arity(*this); }
+int64_t DDim::size() const { return arity(*this); }
 
 bool DDim::operator==(DDim d) const {
   if (var.which() != d.getVar().which()) {
     return false;
   } else {
-    std::vector<int> v1 = vectorize(*this);
-    std::vector<int> v2 = vectorize(d);
+    std::vector<int64_t> v1 = vectorize(*this);
+    std::vector<int64_t> v2 = vectorize(d);
 
     for (unsigned int i = 0; i < v1.size(); i++) {
       if (v1[i] != v2[i]) {
@@ -139,10 +139,10 @@ bool DDim::operator==(DDim d) const {
 bool DDim::operator!=(DDim d) const { return !(*this == d); }
 
 DDim DDim::operator+(DDim d) const {
-  std::vector<int> v1 = vectorize(*this);
-  std::vector<int> v2 = vectorize(d);
+  std::vector<int64_t> v1 = vectorize(*this);
+  std::vector<int64_t> v2 = vectorize(d);
 
-  std::vector<int> v3;
+  std::vector<int64_t> v3;
 
   assert(v1.size() == v2.size());
 
@@ -154,10 +154,10 @@ DDim DDim::operator+(DDim d) const {
 }
 
 DDim DDim::operator*(DDim d) const {
-  std::vector<int> v1 = vectorize(*this);
-  std::vector<int> v2 = vectorize(d);
+  std::vector<int64_t> v1 = vectorize(*this);
+  std::vector<int64_t> v2 = vectorize(d);
 
-  std::vector<int> v3;
+  std::vector<int64_t> v3;
 
   assert(v1.size() == v2.size());
 
@@ -168,15 +168,15 @@ DDim DDim::operator*(DDim d) const {
   return make_ddim(v3);
 }
 
-int get(const DDim& ddim, int idx) { return ddim[idx]; }
+int64_t get(const DDim& ddim, int idx) { return ddim[idx]; }
 
 void set(DDim& ddim, int idx, int value) { ddim[idx] = value; }
 
 /// @cond HIDDEN
 struct VectorizeVisitor : public boost::static_visitor<> {
-  std::vector<int>& vector;
+  std::vector<int64_t>& vector;
 
-  explicit VectorizeVisitor(std::vector<int>& v) : vector(v) {}
+  explicit VectorizeVisitor(std::vector<int64_t>& v) : vector(v) {}
 
   template <typename T>
   void operator()(const T& t) {
@@ -188,31 +188,31 @@ struct VectorizeVisitor : public boost::static_visitor<> {
 };
 /// @endcond
 
-std::vector<int> vectorize(const DDim& ddim) {
-  std::vector<int> result;
+std::vector<int64_t> vectorize(const DDim& ddim) {
+  std::vector<int64_t> result;
   VectorizeVisitor visitor(result);
   boost::apply_visitor(visitor, ddim);
   return result;
 }
 
-struct ProductVisitor : public boost::static_visitor<ssize_t> {
+struct ProductVisitor : public boost::static_visitor<int64_t> {
   template <int D>
-  ssize_t operator()(const Dim<D>& dim) {
+  int64_t operator()(const Dim<D>& dim) {
     return product(dim);
   }
 };
 
-ssize_t product(const DDim& ddim) {
+int64_t product(const DDim& ddim) {
   ProductVisitor visitor;
   return boost::apply_visitor(visitor, ddim);
 }
 
 struct SliceVectorizeVisitor : public boost::static_visitor<> {
-  std::vector<int>& vector;
+  std::vector<int64_t>& vector;
   int begin;
   int end;
 
-  SliceVectorizeVisitor(std::vector<int>& v, int b, int e)
+  SliceVectorizeVisitor(std::vector<int64_t>& v, int b, int e)
       : vector(v), begin(b), end(e) {
     PADDLE_ENFORCE(begin < end,
                    "Begin index must be less than end index in ddim slice.");
@@ -240,7 +240,7 @@ struct SliceVectorizeVisitor : public boost::static_visitor<> {
 };
 
 DDim slice_ddim(const DDim& dim, int begin, int end) {
-  std::vector<int> vec;
+  std::vector<int64_t> vec;
   vec.reserve(end - begin);
   SliceVectorizeVisitor visitor(vec, begin, end);
   boost::apply_visitor(visitor, dim);
@@ -280,7 +280,7 @@ std::ostream& operator<<(std::ostream& os, const DDim& ddim) {
   return os;
 }
 
-DDim::DDim(std::initializer_list<int> init_list) {
+DDim::DDim(std::initializer_list<int64_t> init_list) {
   *this = make_ddim(init_list);
 }
 }  // namespace framework

paddle/framework/ddim.h

@@ -40,7 +40,7 @@ struct DDim {
   template <int D>
   explicit DDim(const Dim<D>& in) : var(in) {}
 
-  /*implicit*/ DDim(std::initializer_list<int> init_list);
+  /*implicit*/ DDim(std::initializer_list<int64_t> init_list);
 
   template <int D>
   DDim& operator=(const Dim<D>& in) {
@@ -48,8 +48,8 @@ struct DDim {
     return *this;
   }
 
-  int& operator[](int idx);
-  int operator[](int idx) const;
+  int64_t& operator[](int idx);
+  int64_t operator[](int idx) const;
 
   template <typename Visitor>
   typename Visitor::result_type apply_visitor(Visitor& visitor) {
@@ -71,15 +71,15 @@ struct DDim {
 
   DDim operator*(DDim d) const;
 
-  ssize_t size() const;
+  int64_t size() const;
 };
 
 /**
- * \brief Make a DDim from std::vector<int>
+ * \brief Make a DDim from std::vector<int64_t>
  *
  * \param dims An vector of ints. Must be sized between [1, 9]
  */
-DDim make_ddim(const std::vector<int>& dims);
+DDim make_ddim(const std::vector<int64_t>& dims);
 
 /**
  * \brief Make a DDim from an initializer list
@@ -87,14 +87,14 @@ DDim make_ddim(const std::vector<int>& dims);
  * \param dims An initializer list of ints. Must be sized between [1, 9]
  *
  */
-DDim make_ddim(std::initializer_list<int> dims);
+DDim make_ddim(std::initializer_list<int64_t> dims);
 
-int get(const DDim& dim, int idx);
+int64_t get(const DDim& dim, int idx);
 void set(DDim& dim, int idx, int val);
 
-std::vector<int> vectorize(const DDim& ddim);
+std::vector<int64_t> vectorize(const DDim& ddim);
 
-ssize_t product(const DDim& ddim);
+int64_t product(const DDim& ddim);
 
 /**
  * \brief Slice a ddim

paddle/framework/ddim_test.cc

@@ -12,7 +12,7 @@ TEST(DDim, Equality) {
   EXPECT_EQ(ddim[2], 5);
 
   // construct a DDim from a vector
-  std::vector<int> vec({9, 1, 5});
+  std::vector<int64_t> vec({9, 1, 5});
   paddle::framework::DDim vddim = paddle::framework::make_ddim(vec);
   EXPECT_EQ(ddim[0], 9);
   EXPECT_EQ(ddim[1], 1);
@@ -25,7 +25,7 @@ TEST(DDim, Equality) {
   EXPECT_EQ(paddle::framework::get(ddim, 0), 6);
 
   // vectorize a DDim
-  std::vector<int> res_vec = paddle::framework::vectorize(vddim);
+  std::vector<int64_t> res_vec = paddle::framework::vectorize(vddim);
   EXPECT_EQ(res_vec[0], 9);
   EXPECT_EQ(res_vec[1], 1);
   EXPECT_EQ(res_vec[2], 5);

paddle/framework/dim.h

@@ -17,13 +17,13 @@ struct Dim {
   static constexpr int dimensions = i;
 
   template <typename... Args>
-  HOSTDEVICE Dim(int _head, Args... _tail) : head(_head), tail(_tail...) {
+  HOSTDEVICE Dim(int64_t _head, Args... _tail) : head(_head), tail(_tail...) {
     static_assert(sizeof...(_tail) == i - 1,
                   "Dim initialized with the wrong number of parameters");
   }
 
   HOSTDEVICE
-  Dim(int _head, const Dim<i - 1>& _tail) : head(_head), tail(_tail) {}
+  Dim(int64_t _head, const Dim<i - 1>& _tail) : head(_head), tail(_tail) {}
 
   HOSTDEVICE
   Dim() : head(0), tail() {}
@@ -31,12 +31,12 @@ struct Dim {
   /** Construct a Dim from a linear index and size.  Uses Fortran order
    * indexing. */
   HOSTDEVICE
-  Dim(int idx, const Dim<i>& size)
+  Dim(int64_t idx, const Dim<i>& size)
       : head(idx % size.head), tail(idx / size.head, size.tail) {}
 
   /** Construct a Dim with each dimension set to the given index */
   HOSTDEVICE
-  Dim(int idx) : head(idx), tail(idx) {}
+  Dim(int64_t idx) : head(idx), tail(idx) {}
 
   HOSTDEVICE
   bool operator==(const Dim<i>& o) const {
@@ -47,13 +47,13 @@ struct Dim {
   bool operator!=(const Dim<i>& o) const { return !(*this == o); }
 
   HOSTDEVICE
-  int& operator[](int idx);
+  int64_t& operator[](int idx);
   HOSTDEVICE
-  int operator[](int idx) const;
+  int64_t operator[](int idx) const;
 
   HOST std::string to_string() const;
 
-  int head;
+  int64_t head;
   Dim<i - 1> tail;
 };
 
@@ -63,7 +63,7 @@ struct Dim<1> {
   static constexpr int dimensions = 1;
 
   HOSTDEVICE
-  Dim(int _head) : head(_head) {}
+  Dim(int64_t _head) : head(_head) {}
 
   HOSTDEVICE
   Dim() : head(0) {}
@@ -86,11 +86,11 @@ struct Dim<1> {
   bool operator!=(const Dim<1>& o) const { return !(*this == o); }
 
   HOSTDEVICE
-  int& operator[](int idx);
+  int64_t& operator[](int idx);
   HOSTDEVICE
-  int operator[](int idx) const;
+  int64_t operator[](int idx) const;
 
-  int head;
+  int64_t head;
 };
 
 namespace {
@@ -100,12 +100,12 @@ template <int i>
 struct DimGetter {
   // Return a copy if Dim is const
   template <typename D>
-  HOSTDEVICE static int impl(const D& d) {
+  HOSTDEVICE static int64_t impl(const D& d) {
     return DimGetter<i - 1>::impl(d.tail);
   }
   // Return a reference if Dim is mutable
   template <typename D>
-  HOSTDEVICE static int& impl(D& d) {
+  HOSTDEVICE static int64_t& impl(D& d) {
     return DimGetter<i - 1>::impl(d.tail);
   }
 };
@@ -115,18 +115,18 @@ template <>
 struct DimGetter<0> {
   // Return a copy if Dim is const
   template <typename D>
-  HOSTDEVICE static int impl(const D& d) {
+  HOSTDEVICE static int64_t impl(const D& d) {
     return d.head;
   }
   // Return a reference if Dim is mutable
   template <typename D>
-  HOSTDEVICE static int& impl(D& d) {
+  HOSTDEVICE static int64_t& impl(D& d) {
     return d.head;
   }
 };
 
 template <int D>
-HOSTDEVICE int& indexer(Dim<D>& dim, int idx) {
+HOSTDEVICE int64_t& indexer(Dim<D>& dim, int idx) {
 #ifndef __CUDA_ARCH__
   if (idx < 0) {
     throw std::invalid_argument("Tried to access a negative dimension");
@@ -141,7 +141,7 @@ HOSTDEVICE int& indexer(Dim<D>& dim, int idx) {
 }
 
 template <>
-HOSTDEVICE int& indexer<1>(Dim<1>& dim, int idx) {
+HOSTDEVICE int64_t& indexer<1>(Dim<1>& dim, int idx) {
 #ifndef __CUDA_ARCH__
   if (idx != 0) {
     throw std::invalid_argument("Invalid index");
@@ -153,7 +153,7 @@ HOSTDEVICE int& indexer<1>(Dim<1>& dim, int idx) {
 }
 
 template <int D>
-HOSTDEVICE int indexer(const Dim<D>& dim, int idx) {
+HOSTDEVICE int64_t indexer(const Dim<D>& dim, int idx) {
 #ifndef __CUDA_ARCH__
   if (idx < 0) {
     throw std::invalid_argument("Tried to access a negative dimension");
@@ -168,7 +168,7 @@ HOSTDEVICE int indexer(const Dim<D>& dim, int idx) {
 }
 
 template <>
-HOSTDEVICE int indexer<1>(const Dim<1>& dim, int idx) {
+HOSTDEVICE int64_t indexer<1>(const Dim<1>& dim, int idx) {
 #ifndef __CUDA_ARCH__
   if (idx != 0) {
     throw std::invalid_argument("Invalid index");
@@ -182,73 +182,76 @@ HOSTDEVICE int indexer<1>(const Dim<1>& dim, int idx) {
 }  // namespace
 // Static access to constant Dim
 template <int i, int l>
-HOSTDEVICE int get(const Dim<l>& d) {
+HOSTDEVICE int64_t get(const Dim<l>& d) {
   return DimGetter<i>::impl(d);
 }
 
 // Static access to mutable Dim
 template <int i, int l>
-HOSTDEVICE int& get(Dim<l>& d) {
+HOSTDEVICE int64_t& get(Dim<l>& d) {
   return DimGetter<i>::impl(d);
 }
 
 // Dynamic access to constant Dim
 template <int l>
-HOSTDEVICE int Dim<l>::operator[](int i) const {
+HOSTDEVICE int64_t Dim<l>::operator[](int i) const {
   return indexer(*this, i);
 }
 
 // Dynamic access to mutable Dim
 template <int l>
-HOSTDEVICE int& Dim<l>::operator[](int i) {
+HOSTDEVICE int64_t& Dim<l>::operator[](int i) {
   return indexer(*this, i);
 }
 
 // Dynamic access to constant Dim
-inline HOSTDEVICE int Dim<1>::operator[](int i) const {
+inline HOSTDEVICE int64_t Dim<1>::operator[](int i) const {
   return indexer(*this, i);
 }
 
 // Dynamic access to mutable Dim
-inline HOSTDEVICE int& Dim<1>::operator[](int i) { return indexer(*this, i); }
+inline HOSTDEVICE int64_t& Dim<1>::operator[](int i) {
+  return indexer(*this, i);
+}
 
 // Dynamic access to constant Dim
 // without std::enable_if will try to instantiate this on get<0>(d)
 template <int l>
-HOSTDEVICE typename std::enable_if<(l > 0), int>::type get(const Dim<l>& d,
-                                                           int i) {
+HOSTDEVICE typename std::enable_if<(l > 0), int64_t>::type get(const Dim<l>& d,
+                                                               int i) {
   return d[i];
 }
 
 // Dynamic access to mutable Dim
 template <int l>
-HOSTDEVICE typename std::enable_if<(l > 0), int&>::type get(Dim<l>& d, int i) {
+HOSTDEVICE typename std::enable_if<(l > 0), int64_t&>::type get(Dim<l>& d,
+                                                                int i) {
   return d[i];
 }
 
 // Dot product of two dims
 template <int i>
-HOSTDEVICE int linearize(const Dim<i>& a, const Dim<i>& b) {
+HOSTDEVICE int64_t linearize(const Dim<i>& a, const Dim<i>& b) {
   return a.head * b.head + linearize(a.tail, b.tail);
 }
 
 // Base case dot product of two Dims
 // Notice it is inline because it is no longer a template
 template <>
-HOSTDEVICE inline int linearize(const Dim<1>& a, const Dim<1>& b) {
+HOSTDEVICE inline int64_t linearize(const Dim<1>& a, const Dim<1>& b) {
   return a.head * b.head;
 }
 
 // Product of a Dim
 template <int i>
-HOSTDEVICE int product(const Dim<i>& a, int prod = 1) {
+HOSTDEVICE int64_t product(const Dim<i>& a, int prod = 1) {
   return prod * a.head * product(a.tail);
 }
 
 // Base case product of a Dim
 // Notice it is inline because it is no longer a template
 template <>
-HOSTDEVICE inline int product(const Dim<1>& a, int prod) {
+HOSTDEVICE inline int64_t product(const Dim<1>& a, int prod) {
   return prod * a.head;
 }
 

paddle/framework/dim_test.cu

@@ -8,7 +8,7 @@ __global__ void test(paddle::framework::Dim<2>* o) {
   o[0] = paddle::framework::make_dim(5, 6);
 }
 
-__global__ void dyn_idx_gpu(int* o) {
+__global__ void dyn_idx_gpu(int64_t* o) {
   auto d = paddle::framework::make_dim(5, 6);
   o[0] = d[1];
 }
@@ -47,9 +47,9 @@ TEST(Dim, Equality) {
   EXPECT_EQ(b[1], 11);
 
   // dynamic access on GPU
-  thrust::device_vector<int> r(1);
+  thrust::device_vector<int64_t> r(1);
   dyn_idx_gpu<<<1, 1>>>(thrust::raw_pointer_cast(r.data()));
-  int res = r[0];
+  int64_t res = r[0];
   EXPECT_EQ(res, 6);
 
   // ex_prefix_mul