Merge remote-tracking branch 'upstream/develop' into mkl_packed

CMakeLists.txt

@@ -16,8 +16,6 @@ cmake_minimum_required(VERSION 3.0)
 set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} "${CMAKE_CURRENT_SOURCE_DIR}/cmake")
 set(PADDLE_SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR})
 set(PADDLE_BINARY_DIR ${CMAKE_CURRENT_BINARY_DIR})
-SET(CMAKE_CXX_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG")
-SET(CMAKE_C_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG")
 
 include(system)
 
@@ -201,6 +199,10 @@ if(WITH_GOLANG)
 endif(WITH_GOLANG)
 
 set(PADDLE_PYTHON_BUILD_DIR "${CMAKE_CURRENT_BINARY_DIR}/python/build")
+
+SET(CMAKE_CXX_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG")
+SET(CMAKE_C_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG")
+
 add_subdirectory(paddle)
 if(WITH_PYTHON)
   add_subdirectory(python)

benchmark/IntelOptimizedPaddle.md

@@ -22,6 +22,7 @@ On each machine, we will test and compare the performance of training on single
 
 #### Training
 Test on batch size 64, 128, 256 on Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz
+Pay attetion that the speed below includes forward, backward and parameter update time. So we can not directly compare the data with the benchmark of caffe `time` [command](https://github.com/PaddlePaddle/Paddle/blob/develop/benchmark/caffe/image/run.sh#L9), which only contain forward and backward. The updating time of parameter would become very heavy when the weight size are large, especially on alexnet.
 
 Input image size - 3 * 224 * 224, Time: images/second
 
@@ -55,6 +56,16 @@ Input image size - 3 * 224 * 224, Time: images/second
 
 <img src="figs/googlenet-cpu-train.png" width="500">
 
+- Alexnet
+
+| BatchSize    | 64     | 128    | 256    |
+|--------------|--------| ------ | -------|
+| OpenBLAS     | 2.13   | 2.45   | 2.68   | 
+| MKLML        | 66.37  | 105.60 | 144.04 |
+| MKL-DNN      | 399.00 | 498.94 | 626.53 | 
+
+chart TBD
+
 #### Inference
 Test on batch size 1, 2, 4, 8, 16 on Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz
 - VGG-19
@@ -82,6 +93,15 @@ Test on batch size 1, 2, 4, 8, 16 on Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz
 | MKLML     | 22.74  | 41.56  | 81.22  | 133.47 | 210.53 |
 | MKL-DNN   | 175.10 | 272.92 | 450.70 | 512.00 | 600.94 |
 
+- Alexnet
+
+| BatchSize | 1      | 2      | 4      | 8      | 16     |
+|-----------|--------|--------|--------|--------|--------|
+| OpenBLAS  |    |   |   |   |    |
+| MKLML     | 21.32  | 36.55  | 73.06  | 131.15 | 192.77 |
+| MKL-DNN   | 442.91 | 656.41 | 719.10 | 847.68 | 850.51 |
+
+chart TBD
 
 ### Laptop
 TBD

benchmark/paddle/image/alexnet.py

@@ -19,7 +19,11 @@ args = {
     'num_samples': num_samples
 }
 define_py_data_sources2(
-    "train.list", None, module="provider", obj="process", args=args)
+    "train.list" if not is_infer else None,
+    "test.list" if is_infer else None,
+    module="provider",
+    obj="process",
+    args=args)
 
 settings(
     batch_size=batch_size,

benchmark/paddle/image/run_openblas_infer.sh

@@ -8,15 +8,19 @@ function clock_to_seconds() {
 }
 
 function infer() {
-  unset OMP_NUM_THREADS MKL_NUM_THREADS OMP_DYNAMIC KMP_AFFINITY
   topology=$1
   layer_num=$2
   bs=$3
-  thread=`nproc`
-  if [ $thread -gt $bs ]; then
-    thread=$bs
+  trainers=`nproc`
+  if [ $trainers -gt $bs ]; then
+    trainers=$bs
   fi
-  log="logs/infer-${topology}-${layer_num}-${thread}openblas-${bs}.log"
+  log="logs/infer-${topology}-${layer_num}-${trainers}openblas-${bs}.log"
+  threads=$((`nproc` / trainers))
+  if [ $threads -eq 0 ]; then
+    threads=1
+  fi
+  export OPENBLAS_NUM_THREADS=$threads
 
   models_in="models/${topology}-${layer_num}/pass-00000/"
   if [ ! -d $models_in ]; then
@@ -28,7 +32,7 @@ function infer() {
     --config="${topology}.py" \
     --use_mkldnn=False \
     --use_gpu=False \
-    --trainer_count=$thread \
+    --trainer_count=$trainers \
     --log_period=$log_period \
     --config_args="batch_size=${bs},layer_num=${layer_num},is_infer=True,num_samples=256" \
     --init_model_path=$models_in \

benchmark/paddle/image/run_openblas_train.sh

 set -e
 
 function train() {
-  unset OMP_NUM_THREADS MKL_NUM_THREADS OMP_DYNAMIC KMP_AFFINITY
+  export OPENBLAS_NUM_THREADS=1
   topology=$1
   layer_num=$2
   bs=$3

doc/api/v2/config/layer.rst

@@ -252,6 +252,11 @@ first_seq
 ..  autoclass:: paddle.v2.layer.first_seq
     :noindex:
 
+sub_seq
+---------
+..  autoclass:: paddle.v2.layer.sub_seq
+    :noindex:
+
 concat
 ------
 ..  autoclass:: paddle.v2.layer.concat

doc/api/v2/fluid/layers.rst

@@ -68,12 +68,6 @@ scale
     :noindex:
 
 
-reshape
----------
-..  autofunction:: paddle.v2.fluid.layers.reshape
-    :noindex:
-
-
 transpose
 ---------
 ..  autofunction:: paddle.v2.fluid.layers.transpose
@@ -170,6 +164,18 @@ sequence_pool
     :noindex:
 
 
+sequence_first_step
+-------------------
+..  autofunction:: paddle.v2.fluid.layers.sequence_first_step
+    :noindex:
+
+
+sequence_last_step
+------------------
+..  autofunction:: paddle.v2.fluid.layers.sequence_last_step
+    :noindex:
+
+
 pool2d
 ------
 ..  autofunction:: paddle.v2.fluid.layers.pool2d
@@ -320,7 +326,19 @@ reduce_sum
 
 
 reduce_mean
----------
+-----------
 ..  autofunction:: paddle.v2.fluid.layers.reduce_mean
     :noindex:
 
+
+reduce_max
+----------
+..  autofunction:: paddle.v2.fluid.layers.reduce_max
+    :noindex:
+
+
+reduce_min
+----------
+..  autofunction:: paddle.v2.fluid.layers.reduce_min
+    :noindex:
+

doc/design/backward.md

+# Backward Building
+
+## Motivation
+
+In Neural Network, most models are solved by the backpropagation algorithm(known as **BP**) at present. Technically, BP calculates the gradient of the loss function, then propagates it back through the networks following the chain rule. However, when configuring the model structure, users do not need to define the backward part. So a mechanism is required by the framework which can complete the model's backward part automatically according to the given forward part.
+
+When implementing a specific `op`, the developer is also asked to implement its backward version, called `grad_op`. A `grad_op` takes gradients of its corresponding `op`'s outputs, and calculate gradients of the `op`'s inputs. During the building of a model's backward part, the framework creates each forward `op`'s `grad_op`, and then string them together in reverse order of forwarding part. In this way, gradients spread from the end to the beginning of the model, in another word, from the loss to parameters.
+
+## Challenges
+
+The motivation of backward building is apparent. However, implementation it correctly is not so easy. In the **Fluid** design, a deep learning model is described by `Program`, `Block`, `Op` and `Variable`. The `Block` itself can be nested. It means that the `op`s and `variable`s are scattered across different blocks rather than all be gathered in a single graph. Our backward building algorithm shall visit blocks in recursive order and be able to insert `grad_op`s and new created `variable`s into the right place. 
+
+## Usage
+
+Although the whole algorithm is comprised of many functions, only one is exposed as API:
+
+```python
+def append_backward(loss, parameter_list=None, no_grad_set=None):
+    """
+    Append backward part to main_program
+
+    Args:
+        loss(Variable): The variable generated by the cost function.
+        parameter_list(list): Parameters that need to be updated by optimizers.
+            If None, it means all parameters need to be updated.
+
+        no_grad_set(set): Variables that have no gradients in Block 0. 
+            If None, the set will be generated inside the function and 
+            contains all variables with `step_gradient=True` from all blocks.
+        
+    Return:
+        (list[Variable]): list of (parameters, gradients) pair.
+    """
+```
+
+By invoking this API, the framework appends backward part of the program where the `loss` is. It takes three arguments. `loss` means the final loss value. It must be a scalar and is usually the output of the loss layer. It is also where the gradient generated and backpropagation starts. `parameter_list` marks all parameters needs updating. If it's `None`, all parameter will be updated by optimizers. `no_grad_set` marks variables without gradient. if all outputs of some `grad_op` are in `no_grad_set`, the `grad_op` will not be run.
+
+This API will be invoked automatically before optimizer building. 
+As a result, in most cases, users do not need to invoke the API by themselves to append backward part.
+
+## Implementation
+
+The implementation of backward building algorithm is in `backward.py` file. The whole algorithm can be divided into two independent parts: creating `grad_op`s and creating new variables. 
+
+### Creating `grad_op`s
+
+The creating of `grad_op`s is implemented by:
+
+```python
+def _append_backward_ops_(target,
+                          block,
+                          target_block,
+                          no_grad_dict,
+                          grad_to_var):
+    """
+    Create all grad ops, and insert them into given block
+
+    Args:
+        target(Variable): the target variable of forward pass
+        block(Block): the block where forward ops are
+        target_block(Block): the block which is going to hold new generated grad ops
+        no_grad_dict(dict): 
+            key(int)  block index
+            val(set) a set of varibale names. These varibales have no gradient
+        grad_to_var(dict)(output argument):
+            key(str): grad variable name
+            val(str): corresponding forward variable name
+    """
+```
+
+Given a `block`, the function will traverses all `op`s in this block in reverse order, gets corresponding `grad_op` from the C++ core via `core.get_grad_op_desc()`, then append it to `target_block`. 
+
+However, some specific `op`(e.g. `while_op`, `if_else_op`) can hold its own sub-block. For these sub-blocks contains `op`s as well, the `grad_op` creating should be recursive.
+
+During the reverse traversal, we check each `op` whether it has an attribute named `sub_block`. If so, it means there is a sub-block and we need to deal with it first. After creating a new block whose father is the one in `op`'s attribute, we invoke `_append_backward_ops_()` recursively, assigning the new block to parameter `target_block` and the one in `op`'s attribute to `block`. The *pseudo-code* shows this process:
+
+```
+******* pseudo-code ********
+for op in reversed(block.ops):
+    if op has an attribute named 'sub_block':
+        Get the sub-block(`s_block`) from op's attribute.
+        Create a new block(`grad_s_block`), whose father is `s_block`.
+        Invoke _append_backward_ops_(), with `block=s_block` and `target_block=grad_s_block`
+    
+    Invoke `core.get_grad_op_desc()` to get op's grad_op.
+    Insert name correspondings between variables and their gradients of the grad_op to grad_to_var
+    Assign grad_s_block to grad_op as it's 'sub_block' attribute.
+    Append grad_op to current target_block.
+```
+
+The first invoking of `_append_backward_ops_()` is initiated by `append_backward()`, in which parameters `block` and `target_block` are all assigned with root block(the block with index 0).
+
+### Corner Cases of `grad_op` Creating
+
+In the previous section, we show the regular process of `grad_op` creating. However, in some corner cases, the conventional algorithm is not enough to get the correct result and appending handling is required. These additional processes run after the algorithm mentioned above and do some special adjusts on its output `grad_op`s.
+
+#### Shared Variables
+
+If a variable is read by more than one `op` in the forward pass, its gradient is likely to be written by more than one `grad_op`s in the next backward pass. To make the gradient result being the sum of all `grad_op`s' outputs instead of the last running one, we assign each output with a temporary variable and then add a `sum_op` to add them up. 
+
+For the debug convenience, if the final gradient name is `w@GRAD`, it's corresponding temporary variables will be named as `w@GRAD@RENAME@0`, `w@GRAD@RENAME@1`...
+
+See function `_addup_repetitive_outputs_` in `backward.py` for implementation details.
+
+#### No Gradient Variables
+
+In our framework, variables can be marked as *no_gradient*, it means that the gradient of this variable is unnecessary and can be considered as zero in model training. Apparently, when all the outputs of some `grad_op` are marked as *no_gradient*, the `grad_op` itself can be skipped in backward pass. 
+
+But these unnecessary gradients still need to be creating and initialized by something, otherwise following `grad_op`s who take these gradients as inputs take the risk of using uninitialized memory. In our code, we employ `fill_zeros_like_op` to initialize them as all zeros. 
+
+This features are implemented in function `_remove_no_grad_branch_`. It checks new created `grad_op`s one-by-one, removes whose outputs are all in `no_grad_set` or inserts `fill_zeros_like_op` when its necessary. We can get the `no_grad_set` from the `_append_backward_ops_` argument `no_grad_dict` or generate it on the fly by scanning all variables' `no_gradient` attribute(True or False). 
+
+### Creating Backward Variables
+
+Up to now, we have completed all creating and adjusting jobs of `grad_op`s. However, backward variables have not been created. Now they are only represented by `grad_op`'s input and output arguments. The backward variable creating job will be done by:
+
+```python
+def _append_backward_vars_(block, 
+                           start_op_idx, 
+                           grad_to_var, 
+                           grad_info_map):
+    """
+    Create new variables required by backward pass.
+
+    Args:
+        block(Block): the block where new variables will be created
+        start_op_idx(int): Only variables required by ops in block.ops[start_op_idx : ] will be created
+        grad_to_var(dict):
+            key(str): grad variable name
+            val(str): corresponding forward variable name
+            In most cases, this dict is generated by _append_backward_ops_()
+        grad_info_map(dict)(output argument):
+            key(str): forward variable name
+            val(tuple): a tuple of (str, int), str is the corresponding grad name, int is the block index
+    """
+```
+
+Given a `block`, this function traverses all the `grad_op`s in it(The argument `start_op_idx` indicates where the grad_op sequence starts.) and creates all the uncreated outputs. The *pseudo-code* shows this process:
+
+```
+for op in block.ops[start_op_idx : ]:
+
+    if op has an attribute named 'sub_block':
+        Get the sub-block(`s_block`) from op's attribute.
+        Invoke _append_backward_vars_(), with `block=s_block`
+        
+    for var_name in op.all_output_names():
+        if block.has_var_recursive(var_name) or var_name is the name of empty variable:
+            continue
+        create a new variable named 'var_name' in block
+        if grad_to_var.has_key(var_name):
+            set grad_info_map[grad_to_var[var_name]] as a tuple of (var_name. block)
+            
+    do op's var type inference
+    do op's shape inference
+```

doc/design/block.md

@@ -291,10 +291,10 @@ public:
   }
 
   void Run(const framework::Scope& scope,
-           const platform::DeviceContext& dev_ctx) const override {
+           const platform::Place& place) const override {
     PADDLE_ENFORCE(symbols_ready_, "operators and variables should be created first.");
     for (auto& op : runtime_table_.ops()) {
-      op->Run(scope, dev_ctx);
+      op->Run(scope, place);
     }
   }
 

doc/design/mkl/mkldnn_fluid.md

+# Design Doc: Add MKLDNN Kernel in Fluid Operator
+
+## Principles
+
+First of all, we should follow some basical principles like:
+1.  [How to write a new operator](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/new_op_en.md). We are trying to add a new kind of kernel into operators, so basically we should follow this doc.
+2.  [Supporting new Device/Library](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/support_new_device.md). Since MKLDNN is a new library to fluid, we should add `MKLDNNDeviceContext` and maybe `mkldnn_helper.h`, just like [cudnn_helper.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/cudnn_helper.h).
+3.  [Switch Kernel](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/switch_kernel.md). Another important point is that we should ensure the data synchronization between different kernel types, which is this [topic](https://github.com/PaddlePaddle/Paddle/issues/6549). So basically we should override `GetExpectedKernelType` and `trans` functions to support switching kernels.
+4.  [The Keys of Operator Kernel Type](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/operator_kernel_type.md). Kernel Type is a pivotal conception which can record the `Place`, `Library`, `DataType` and `Layout`.
+
+## Sulution
+
+In general, there are four parts we should follow to run a MKL-DNN primitive.
+-  Create a primitive descriptor that describe this operator
+-  Create a primitive itself by primitive descriptor and the engine
+-  Create all memory buffers that primitive needed
+-  Launch a stream to execute the primitive created
+More details can refer to [here](http://01org.github.io/mkl-dnn).
+
+It's better to avoid reinitialization of primitives and memory handles in the first three stages in every iteration. \
+So we plan to create a map to record all the `primitive` and `memory`, which should not take too much memories as discussed [here](https://github.com/PaddlePaddle/Paddle/issues/6822).
+
+It's assumed that following three conditions should be satisfied.
+1. there is a unique key for each operator instance. May be the actual name of `Output Tensor`.
+2. the `Input Tensor` inside `Compute` function is the one after converted.
+3. we can get the phase(eg. `is_test`) inside `Compute` function, otherwise we need to expose this attribue to user.
+
+### Compute
+The algorithm of `Compute` would be described as follow, let's take conv like an example.
+
+```c++
+
+  PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), "It must use CPUPlace.");
+  PADDLE_ENFORCE(platform::is_mkldnn_library(ctx.GetLibrary()), "It must use MKLDNN Library.");
+
+  auto& dev_ctx = ctx.template device_context<platform::MKLDNNDeviceContext>();
+
+  // find primitive by unique key from mkldnn context
+  // the op_key should be a unique name of this op instance
+  auto& p = dev_ctx.findPrimitive(op_key + "_fwd");
+
+  // assuming the input tensor inside this compute function is the one after converted
+  // this point should be guarantee by another mechanism
+  auto& i = dev_ctx.findMemory(op_key + "_input");
+  
+  if (p == nullptr || i == nullptr || inputSizeChanged(p, i))  {
+    auto fwd_primitive_desc = createPrimitiveDesc(ctx);
+    auto* input = ctx.Input<Tensor>("Input");
+    auto* filter = ctx.Input<Tensor>("Filter");
+    auto* output = ctx.Output<Tensor>("Output");
+    shared_ptr<mkldnn::memory> in(new mkldnn::memory(fwd_primitive_desc->src_primitive_desc(), input->data<T>()));
+    shared_ptr<mkldnn::memory> wgt(new mkldnn::memory(fwd_primitive_desc->weights_primitive_desc(), filter->data<T>()));
+    shared_ptr<mkldnn::memory> out(new mkldnn::memory(fwd_primitive_desc->dst_primitive_desc(), output->mutable_data<T>(ctx.GetPlace())));
+    shared_ptr<mkldnn::conv_fwd> fwd_primitive(new mkldnn::conv_fwd(*fwd_primitive_desc, *in, *wgt, *out));
+
+    dev_ctx.addMemory(op_key+"_input", in);
+    dev_ctx.addMemory(op_key+"_output", out);
+    dev_ctx.addMemory(op_key+"_filer", wgt);
+    dev_ctx.addPrimitive(op_key+"_fwd", fwd_primitive);
+    dev_ctx.addPrimitiveDesc(op_key+"_fwd_PD", fwd_primitive_desc);
+  }
+
+  p = dev_ctx.findPrimitive(op_key + "_fwd");
+
+  PADDLE_ENFORCE(p, "Should have forward Primitive");
+  PADDLE_ENFORCE(dev_ctx.findMemory(op_unique_key+"_input"), "Should have input memory");
+  PADDLE_ENFORCE(dev_ctx.findMemory(op_unique_key+"_output"), "Should have output memory");
+  PADDLE_ENFORCE(dev_ctx.findMemory(op_unique_key+"_filter"), "Should have filter memory");
+  PADDLE_ENFORCE(dev_ctx.findPrimitiveDesc(op_unique_key+"_fwd_PD"), "Should have forward PrimitiveDesc");
+  dev_ctx.submit(p);
+  dev_ctx.execute();  // the convert primitive should have already contained.
+
+```
+
+The `createPrimitiveDesc` returns the primitive descripotor of this operator, would be like this:
+```c++
+  auto* input = ctx.Input<Tensor>("Input");
+  auto* filter = ctx.Input<Tensor>("Filter");
+  auto* output = ctx.Output<Tensor>("Output");
+  std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
+  std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
+  std::vector<int> dilations = ctx.Attr<std::vector<int>>("dilations");
+  int groups = ctx.Attr<int>("groups");
+  algorithm algo = static_cast<algorithm>(ctx.Attr<int>("convolution_algorithm_option"));
+  prop_kind pk = ctx.Attr<bool>("is_test") ? prop_kind::forward_inference : prop_kind::forward_training;
+    
+  auto fwd_desc = mkldnn::conv_fwd::desc(/* all the setting above*/);
+  shared_ptr<mkldnn::conv_fwd::primitive_desc> fwd_primitive_desc(new mkldnn::conv_fwd::primitive_desc(fwd_desc, ctx.getEngine()));
+
+  return fwd_primitive_desc;
+  }
+```
+
+### MKLDNNDeviceContext
+`MKLDNNDeviceContext`, which is very straightforward, should contain some base information like: `stream`, `engine` and the map needed.
+
+
+### mkldnn_helper
+Some functions would be put in `paddle/platform/mkldnn_helper.h`.
+- create MKLDNN memories
+- create MKLDNN primitives
+- error check function
+- etc
+
+
+### Kernel Switch
+We should `reorder` the different Layout from other device or to other device. `GetExpectedKernelType` and `trans` functions can help us to implement it.
+
+`GetExpectedKernelType` should get the context, and this operator can return the best `KernelType`. 
+`trans` would be like this:
+
+```c++
+void trans(inputs, ctx) override {
+  if (NoNeedTrans()) {
+    return;
+  }
+  // find reorder primitive by op_key from context
+  auto& dev_ctx = ctx.template device_context<platform::MKLDNNDeviceContext>();
+  auto& p = dev_ctx.findPrimitive(op_key + "_reorder_input");
+  auto& i = dev_ctx.findMemory(op_key + "_src_input");
+
+  if (p == nullptr || i == nullptr || changeSized(i, input)) {
+    auto prim = createPrimitiveDesc(ctx);
+    auto src = createMemory(memoryDesc(input->dims(), actual_layout), input->data);
+    auto newbuffer = paddle::memory::Alloc(ctx.GetPlace(), input->size_in_bytes());
+    auto dst = createMemory(p->expected_desc(), newbuffer->data);
+    auto reorder_primitive(new mkldnn::reorder(src, dst));
+
+    dev_ctx.addMemory(op_key+"_src_input", src);
+    dev_ctx.addMemory(op_key+"_input", dst);
+    dev_ctx.addPrimitive(op_key+"_reorder_input", reorder_primitive);
+  }
+
+  p = dev_ctx.findPrimitive(op_key + "_reorder_input");
+  PADDLE_ENFORCE(p, "Should have Reorder Primitive");
+  dev_ctx.submit(p);
+  if (! this->isMKLDNNKernel()) {
+    // execute immediately only if this is not mkldnn kernel function.
+    // otherwise, it can be executed with the operator primitive in Compute
+    dev_ctx.stream();
+  }
+  // after submit, the input tensor in ExecutionContext should be changed as the converted one
+  // there should be another mechanism to ensure this
+}
+```
+
+### Unit Test
+All the functions should be tested corresponding.
+TBD

doc/design/optimizer.md

@@ -79,7 +79,7 @@ class Optimizer(object):
     def minimize(self, loss, parameter_list):
         """Add operations to minimize `loss` by updating `parameter_list`.
 
-        This method combines interface `append_backward_ops()` and
+        This method combines interface `append_backward()` and
         `create_optimization_pass()` into one.
         """
         params_grads = self.create_backward_pass(loss, parameter_list)

doc/design/support_new_device.md

@@ -25,13 +25,14 @@ There are mainly three parts that we have to consider while integrating a new de
 
 ### Place and DeviceContext
 
+Please remind that device and computing library are not one-to-one corresponding. A device can have a lot of computing libraries and a computing library can also support several devices.
 
 #### Place
-Fluid uses class [Place](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h#L55) to represent different devices and computing libraries. There are inheritance relationships between different kinds of `Place`.
+Fluid uses class [Place](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h#L55) to represent the device memory where data is located. If we add another device, we have to add corresponding `DevicePlace`.
 
 ```
-        |   CPUPlace   --> MKLDNNPlace
-Place --|   CUDAPlace  --> CUDNNPlace
+        |   CPUPlace
+Place --|   CUDAPlace
         |   FPGAPlace
 ```
 
@@ -43,7 +44,7 @@ typedef boost::variant<CUDAPlace, CPUPlace, FPGAPlace> Place;
 
 #### DeviceContext
 
-Fluid uses class [DeviceContext](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/device_context.h#L30) to manage the resources in different hardwares, such as CUDA stream in `CDUADeviceContext`. There are also inheritance relationships between different kinds of `DeviceContext`.
+Fluid uses class [DeviceContext](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/device_context.h#L30) to manage the resources in different libraries, such as CUDA stream in `CDUADeviceContext`. There are also inheritance relationships between different kinds of `DeviceContext`.
 
 
 ```
@@ -106,7 +107,7 @@ template <typename Place>
 size_t Used(Place place);
 ```
 
-To implementing these interfaces, we have to implement MemoryAllocator for different Devices
+To implement these interfaces, we have to implement MemoryAllocator for different Devices.
 
 
 #### Tensor
@@ -243,6 +244,7 @@ REGISTER_OP_CUDA_KERNEL(
 Generally, we will impelement OpKernel for all Device/Library of an Operator. We can easily train a Convolutional Neural Network in GPU. However, some OpKernel is not sutibale on a specific Device. For example, crf operator can only run on CPU, whereas most other operators can run at GPU. To achieve high performance in such circumstance, we have to switch between different Device/Library.
 
 
-We will discuss how to implement an efficient OpKernel switch policy. 
+For more details, please refer to following docs:
 
-- TBD
+- operator kernel type [doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/operator_kernel_type.md)
+- switch kernel [doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/switch_kernel.md)

doc/faq/build_and_install/index_cn.rst

@@ -109,3 +109,31 @@ PaddlePaddle使用avx SIMD指令提高cpu执行效率，因此错误的使用二
 解决办法是：
 
 * 卸载PaddlePaddle包 :code:`pip uninstall paddle`, 清理掉老旧的PaddlePaddle安装包，使得单元测试有一个干净的环境。如果PaddlePaddle包已经在python的site-packages里面，单元测试会引用site-packages里面的python包，而不是源码目录里 :code:`/python` 目录下的python包。同时，即便设置 :code:`PYTHONPATH` 到 :code:`/python` 也没用，因为python的搜索路径是优先已经安装的python包。
+
+8. 下载MKLML库失败
+------------------
+
+..  code-block:: bash
+
+    make[2]: *** [third_party/mklml/src/extern_mklml-stamp/extern_mklml-download] 错误 4
+    make[1]: *** [CMakeFiles/extern_mklml.dir/all] 错误 2
+    make[1]: *** 正在等待未完成的任务....
+
+原因：网速或SSL链接原因，导致MKLML库下载不成功。
+
+解决办法是：手动下载并安装，具体步骤如下。
+
+..  code-block:: bash
+
+    // 1. 进入对应的目录
+    cd build/third_party/mklml/src/extern_mklml
+
+    // 2. 查看包的大小， 正常情况下是75M，如果小于75M，即下载失败：
+    du -sh mklml_lnx_2018.0.1.20171007.tgz
+
+    // 3. 手动下载且解压缩，并手动生成download成功标签：
+    wget --no-check-certificate https://github.com/01org/mkl-dnn/releases/download/v0.11/mklml_lnx_2018.0.1.20171007.tgz -c -O mklml_lnx_2018.0.1.20171007.tgz 
+    tar zxf mklml_lnx_2018.0.1.20171007.tgz
+    touch ../extern_mklml-stamp/extern_mklml-download
+
+    // 4. 接着编译即可

doc/getstarted/build_and_install/docker_install_cn.rst

@@ -15,7 +15,7 @@
 获取PaddlePaddle的Docker镜像
 ------------------------------
 
-执行下面的命令获取最新的PaddlePaddle Docker镜像
+执行下面的命令获取最新的PaddlePaddle Docker镜像，版本为cpu_avx_mkl：
 
   .. code-block:: bash
 
@@ -27,7 +27,7 @@
 
      docker pull docker.paddlepaddle.org/paddle
 
-下载GPU版本的Docker镜像：
+下载GPU版本（cuda8.0_cudnn5_avx_mkl）的Docker镜像：
 
   .. code-block:: bash
 
@@ -54,7 +54,7 @@
 .. _docker_run:
 
 在Docker中执行PaddlePaddle训练程序
-------------------------------
+----------------------------------
 
 假设您已经在当前目录（比如在/home/work）编写了一个PaddlePaddle的程序 :code:`train.py` （可以参考
 `PaddlePaddleBook <http://www.paddlepaddle.org/docs/develop/book/01.fit_a_line/index.cn.html>`_ 
@@ -82,7 +82,7 @@
 .. _docker_run_book:
 
 使用Docker启动PaddlePaddle Book教程
-------------------------------
+-----------------------------------
 
 使用Docker可以快速在本地启动一个包含了PaddlePaddle官方Book教程的Jupyter Notebook，可以通过网页浏览。
 PaddlePaddle Book是为用户和开发者制作的一个交互式的Jupyter Notebook。

doc/getstarted/build_and_install/docker_install_en.rst

@@ -16,7 +16,7 @@ After you've read above tutorials you may proceed the following steps.
 Pull PaddlePaddle Docker Image
 ------------------------------
 
-Run the following command to download the latest Docker images:
+Run the following command to download the latest Docker images, the version is cpu_avx_mkl:
 
   .. code-block:: bash
 
@@ -28,7 +28,7 @@ For users in China, we provide a faster mirror:
 
      docker pull docker.paddlepaddle.org/paddle
 
-Download GPU version images:
+Download GPU version (cuda8.0_cudnn5_avx_mkl) images:
 
   .. code-block:: bash
 
@@ -58,7 +58,7 @@ and run:
 .. _docker_run:
 
 Launch your training program in Docker
-------------------------------
+--------------------------------------
 
 Assume that you have already written a PaddlePaddle program
 named :code:`train.py` under directory :code:`/home/work` (refer to 

doc/getstarted/build_and_install/pip_install_cn.rst

@@ -11,14 +11,14 @@ PaddlePaddle可以使用常用的Python包管理工具
 ------------------------------
 
 
-执行下面的命令即可在当前机器上安装PaddlePaddle的运行时环境，并自动下载安装依赖软件。
+执行下面的命令即可在当前机器上安装PaddlePaddle的运行时环境，并自动下载安装依赖软件，版本为cpu_avx_openblas。
 
   .. code-block:: bash
 
      pip install paddlepaddle
 
 
-如果需要安装支持GPU的版本，需要执行：
+如果需要安装支持GPU的版本（cuda7.5_cudnn5_avx_openblas），需要执行：
 
   .. code-block:: bash
 

doc/getstarted/build_and_install/pip_install_en.rst

@@ -12,14 +12,14 @@ Install Using pip
 ------------------------------
 
 Run the following command to install PaddlePaddle on the current
-machine, it will also download requirements.
+machine, it will also download requirements, the version is cpu_avx_openblas.
 
   .. code-block:: bash
 
      pip install paddlepaddle
 
 
-If you wish to install GPU version, just run:
+If you wish to install GPU version (cuda7.5_cudnn5_avx_openblas), just run:
 
   .. code-block:: bash
 

doc/getstarted/index_cn.rst

@@ -7,13 +7,13 @@
 ++++++++
 
 PaddlePaddle支持使用pip快速安装，目前支持CentOS 6以上, Ubuntu 14.04以及MacOS 10.12，并安装有Python2.7。
-执行下面的命令完成快速安装：
+执行下面的命令完成快速安装，版本为cpu_avx_openblas：
 
   .. code-block:: bash
 
      pip install paddlepaddle
 
-如果需要安装支持GPU的版本，需要执行：
+如果需要安装支持GPU的版本（cuda7.5_cudnn5_avx_openblas），需要执行：
 
   .. code-block:: bash
 

doc/getstarted/index_en.rst

@@ -8,13 +8,13 @@ Quick Install
 
 You can use pip to install PaddlePaddle with a single command, supports
 CentOS 6 above, Ubuntu 14.04 above or MacOS 10.12, with Python 2.7 installed.
-Simply run the following command to install:
+Simply run the following command to install, the version is cpu_avx_openblas:
 
   .. code-block:: bash
 
      pip install paddlepaddle
 
-If you need to install GPU version, run:
+If you need to install GPU version (cuda7.5_cudnn5_avx_openblas), run:
 
   .. code-block:: bash
 

go/pserver/client/c/test/test_cclient.c

 /* 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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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 <stdio.h>
 #include <stdlib.h>

paddle/framework/CMakeLists.txt

@@ -5,10 +5,18 @@ cc_library(ddim SRCS ddim.cc DEPS eigen3)
 cc_test(ddim_test SRCS ddim_test.cc DEPS ddim)
 nv_test(dim_test SRCS dim_test.cu DEPS ddim)
 
-cc_library(tensor SRCS tensor.cc DEPS ddim place paddle_memory device_context)
+if (WITH_GPU)
+  nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS ddim place paddle_memory device_context framework_proto)
+else()
+  cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS ddim place paddle_memory device_context framework_proto)
+endif ()
 
 cc_test(tensor_test SRCS tensor_test.cc DEPS tensor)
-cc_test(tensor_util_test SRCS tensor_util_test.cc DEPS tensor)
+if (WITH_GPU)
+  nv_test(tensor_util_test SRCS tensor_util_test.cc tensor_util_test.cu DEPS tensor)
+else()
+  cc_test(tensor_util_test SRCS tensor_util_test.cc DEPS tensor)
+endif()
 
 cc_test(eigen_test SRCS eigen_test.cc DEPS tensor)
 
@@ -21,6 +29,8 @@ cc_test(variable_test SRCS variable_test.cc)
 cc_library(scope SRCS scope.cc DEPS glog)
 cc_test(scope_test SRCS scope_test.cc DEPS scope)
 
+cc_library(data_transform SRCS data_transform.cc DEPS math_function tensor framework_proto)
+cc_test(data_transform_test SRCS data_transform_test.cc DEPS data_transform device_context)
 
 cc_library(attribute SRCS attribute.cc DEPS framework_proto)
 cc_test(program_desc_test SRCS program_desc_test.cc DEPS proto_desc
@@ -29,12 +39,13 @@ cc_library(op_proto_maker SRCS op_proto_maker.cc DEPS framework_proto attribute)
 cc_test(op_proto_maker_test SRCS op_proto_maker_test.cc DEPS op_proto_maker)
 cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
 cc_library(shape_inference SRCS shape_inference.cc DEPS ddim attribute)
-cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog shape_inference)
-cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry)
+cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog
+    shape_inference data_transform)
+cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry init)
 cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog)
 
 cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc)
-cc_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
+nv_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
 
 py_proto_compile(framework_py_proto SRCS framework.proto)
 # Generate an empty __init__.py to make framework_py_proto as a valid python module.
@@ -59,5 +70,9 @@ cc_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry
 cc_library(selected_rows SRCS selected_rows.cc DEPS tensor)
 cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows)
 
-cc_library(init SRCS init.cc DEPS gflags executor place stringpiece)
+cc_library(threadpool SRCS threadpool.cc)
+cc_test(threadpool_test SRCS threadpool_test.cc DEPS threadpool)
+cc_library(init SRCS init.cc DEPS gflags device_context place stringpiece)
 cc_test(init_test SRCS init_test.cc DEPS init)
+
+cc_test(op_kernel_type_test SRCS op_kernel_type_test.cc DEPS place device_context framework_proto)

paddle/framework/backward.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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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/backward.h"
 #include "paddle/operators/net_op.h"

paddle/framework/backward.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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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
 

paddle/framework/backward.md

-# Operator/expression 's Backward
-
-## Motivation
-
-In Neural Network, most models are solved by the backpropagation algorithm(known as **BP**) at present. Technically, BP calculates the gradient of the loss function, then propagates it back through the networks following the chain rule. Hence we need a module that chains the gradient operators/expressions together to construct the backward pass. Every forward network needs a backward network to construct the full computation graph. The operator/expression's backward pass will be generated with respect to the forward pass. 
-
-## Implementation
-
-In this design doc, we exported only one API for generating the backward pass.
-
-```c++
-std::unique_ptr<OperatorBase> Backward(const OperatorBase& forwardOp,
-    const std::unordered_set<std::string>& no_grad_vars);
-```
-
-The implementation behind it can be divided into two parts, **Backward Operator Creating** and **Backward Operator Building**.
-
-### Backward Operator Registry
-
-A backward network is built up with several backward operators. Backward operators take forward operators' inputs, outputs, and output gradients and then calculate its input gradients.
-
-|                        | forward operator | backward operator 
-| ---------------------- | ---------------- |------------------------- |		
-| **Operator::inputs_**  | Inputs       | Inputs, Outputs, OutputGradients |	
-| **Operator::outputs_** | Outputs          | InputGradients            |
-
- In most cases, there is a one-to-one relation between the forward and backward operators. These relations are recorded by a global hash map(`OpInfoMap`). To follow the philosophy of minimum core and to make operators pluggable, the registry mechanism is introduced.
-
-For example, we have `mul_op`, and we can register its information and corresponding backward operator by the following macro:
-
-```cpp
-REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad);
-```
-
-`mul` is the operator's type. `MulOp` and `MulOpMaker` are the operator class and the operator maker class respectively.
-
-`mul_grad` is the type of backward operator, and `MulOpGrad` is its class name.
-
-### Backward Opeartor Creating
-
-Given a certain forward operator, we can get its corresponding backward operator by calling:
-
-```cpp
-OperatorBase* bwd_op = BuildGradOp(const OperatorBase* fwd_op);
-```
-
-The function `BuildGradOp` will sequentially execute following processes:
-
-1. Get the `type_` of given forward operator, and then get the corresponding backward operator's type by looking up the `OpInfoMap`.
-
-2. Build two maps named `inputs` and `outputs` to temporarily store backward operator's inputs and outputs. Copy forward operator's `inputs_` and `outputs_` to map `inputs`, except these, are not necessary for gradient computing.
-
-3. Add forward inputs' gradient variables into map `output`, adding forward outputs' gradient variables into map `input`.
-
-4. Building backward operator with `inputs`, `outputs` and forward operator's attributes.
-
-### Backward Network Building
-
-A backward network is a series of backward operators. The main idea of building a backward network is creating backward operators in the inverted sequence and appending them together one by one. There are some corner cases that need special processing.
-
-1. Op 
-
-   When the input forward network is an Op, return its gradient Operator immediately. If all of its outputs are in no gradient set, then return a special `NOP`.
-
-2. NetOp 
-
-   In our design, the network itself is also a kind of operator(**NetOp**). So the operators contained by a big network may be some small network. When the input forward network is a NetOp, it needs to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to the forward NetOp.
-
-3. RnnOp
-
-   RnnOp is a nested stepnet operator.  Backward module needs to recusively call `Backward` for every stepnet.
-
-4. Sharing Variables
-
-   As illustrated in the figure 1 and figure 2, two operators share the same variable name **W@GRAD**, which will overwrite their shared input variable. 
-
-<p align="center">
-<img src="./images/duplicate_op.png" width="50%" ><br/>
-
-​	Figure 1. Sharing variables in operators. 
-
-</p>
-
-​	Sharing variable between operators or same input variable used in multiple operators can lead to duplicate gradient variables. As illustrated in figure 2, we need to rename the gradient names recursively and add a generic add operator to prevent overwriting. 
-
-<p align="center">
-<img src="images/duplicate_op2.png" width="40%" ><br/>
-
-​	Figure 2. Replace sharing variable's gradient with `Add` operator.
-
-</p>
-
-​	Because the framework finds variables according to their names, we need to rename the output links. We add an integer suffix to represent its position in the clockwise direction. 
-
-5. Part of the Gradient is Zero.
-
-   In the whole graph, there is some case of that one operator's gradient is not needed, but its input's gradient is a dependency link of other operator,  we need to fill a same shape gradient matrix in the position. In our implementation, we insert a special `fillZeroLike` operator.
-
-
-Follow these rules above, then collect the sub graph `OutputGradients`/`InputGradients` as the NetOp's and return it.

paddle/framework/backward_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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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/backward.h"
 

paddle/framework/data_layout.h

@@ -13,11 +13,15 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include "paddle/platform/enforce.h"
+
+#include <iostream>
+#include "paddle/platform/enforce.h"
 
 namespace paddle {
 namespace framework {
 
-enum DataLayout {
+enum class DataLayout {
   kNHWC = 0,
   kNCHW = 1,
   kAnyLayout = 2,
@@ -33,5 +37,23 @@ inline DataLayout StringToDataLayout(const std::string& str) {
   }
 }
 
+inline std::string DataLayoutToString(const DataLayout& data_layout) {
+  switch (data_layout) {
+    case DataLayout::kNHWC:
+      return "NHWC";
+    case DataLayout::kNCHW:
+      return "NCHW";
+    case DataLayout::kAnyLayout:
+      return "ANY_LAYOUT";
+    default:
+      PADDLE_THROW("unknown DataLayou %d", data_layout);
+  }
+}
+
+inline std::ostream& operator<<(std::ostream& out, DataLayout l) {
+  out << DataLayoutToString(l);
+  return out;
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/data_transform.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/framework/data_transform.h"
+#include "paddle/framework/lod_tensor.h"
+#include "paddle/platform/device_context.h"
+
+namespace paddle {
+namespace framework {
+
+DataTransformFnMap& DataTransformFnMap::Instance() {
+  static DataTransformFnMap data_transform_map;
+  return data_transform_map;
+}
+
+auto KernelFP32 = OpKernelType(proto::DataType::FP32, platform::CPUPlace(),
+                               DataLayout::kNHWC, LibraryType::kPlain);
+
+auto KernelFP64 = OpKernelType(proto::DataType::FP64, platform::CPUPlace(),
+                               DataLayout::kNHWC, LibraryType::kPlain);
+
+auto KernelNHWC = OpKernelType(proto::DataType::FP64, platform::CPUPlace(),
+                               DataLayout::kNHWC, LibraryType::kPlain);
+
+auto KernelNCHW = OpKernelType(proto::DataType::FP64, platform::CPUPlace(),
+                               DataLayout::kNCHW, LibraryType::kPlain);
+
+void TransDataType(const platform::DeviceContext* ctx,
+                   const KernelTypePair& kernel_pair, const Variable& in,
+                   Variable* out) {
+  PADDLE_ENFORCE(in.IsType<Tensor>(), "Only Support Tensor transform!.");
+  PADDLE_ENFORCE(
+      platform::places_are_same_class(kernel_pair.first.place_,
+                                      kernel_pair.second.place_),
+      "TransDataType Only Support DataType transform on same place!");
+
+  auto src = in.Get<Tensor>();
+  auto* dst = out->GetMutable<Tensor>();
+
+  auto dims = src.dims();
+  dst->Resize(dims);
+  auto dst_type = kernel_pair.second.data_type_;
+  auto src_type = kernel_pair.first.data_type_;
+
+  switch (src_type) {
+    case proto::DataType::FP32:
+      framework::VisitDataType(dst_type, CastDataType<float>(src, dst, ctx));
+      break;
+    case proto::DataType::FP64:
+      framework::VisitDataType(dst_type, CastDataType<double>(src, dst, ctx));
+      break;
+    case proto::DataType::INT32:
+      framework::VisitDataType(dst_type, CastDataType<int>(src, dst, ctx));
+      break;
+    case proto::DataType::INT64:
+      framework::VisitDataType(dst_type, CastDataType<int64_t>(src, dst, ctx));
+      break;
+    case proto::DataType::BOOL:
+      framework::VisitDataType(dst_type, CastDataType<bool>(src, dst, ctx));
+      break;
+    default:
+      PADDLE_THROW("Not support type %d", src_type);
+  }
+}
+
+void TransDataLayout(const platform::DeviceContext* ctx,
+                     const KernelTypePair& kernel_pair, const Variable& in,
+                     Variable* out) {
+  PADDLE_ENFORCE(in.IsType<Tensor>(), "Only Support Tensor transform!.");
+  PADDLE_ENFORCE(
+      platform::places_are_same_class(kernel_pair.first.place_,
+                                      kernel_pair.second.place_),
+      "TransDataType Only Support DataType transform on same place!");
+
+  auto src = in.Get<Tensor>();
+  auto* dst = out->GetMutable<Tensor>();
+  PADDLE_ENFORCE(arity(src.dims()) == 4, "Input Arity Only Suppport 4!");
+
+  auto src_dim = src.dims();
+  dst->Resize(src_dim);
+  auto place = kernel_pair.second.place_;
+  CopyFrom(src, place, *ctx, dst);
+  const std::vector<int> axis = {0, 2, 3, 1};
+
+  std::vector<int64_t> dst_dim;
+  dst_dim.resize(axis.size());
+  for (size_t i = 0; i < axis.size(); i++) {
+    dst_dim[i] = src_dim[axis[i]];
+  }
+
+  dst->Resize(make_ddim(dst_dim));
+
+  auto src_type = kernel_pair.first.data_type_;
+  framework::VisitDataType(src_type, CastDataLayout(src, dst, ctx, axis));
+
+  dst->set_layout(kernel_pair.second.data_layout_);
+}
+
+}  // namespace framework
+}  // namespace paddle
+
+namespace f = paddle::framework;
+REGISTER_DATA_TRANSFORM_FN(f::KernelFP32, f::KernelFP64, f::TransDataType);
+REGISTER_DATA_TRANSFORM_FN(f::KernelNHWC, f::KernelNCHW, f::TransDataLayout);

paddle/framework/data_transform.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 <functional>
+#include <utility>
+#include <vector>
+
+#include "paddle/framework/op_kernel_type.h"
+#include "paddle/framework/tensor.h"
+#include "paddle/framework/variable.h"
+#include "paddle/operators/math/math_function.h"
+#include "paddle/platform/device_context.h"
+#include "paddle/platform/macros.h"
+#include "paddle/platform/transform.h"
+
+namespace paddle {
+namespace framework {
+
+using KernelTypePair = std::pair<OpKernelType, OpKernelType>;
+
+using DataTransformFn =
+    std::function<void(const platform::DeviceContext*, const KernelTypePair&,
+                       const Variable&, Variable*)>;
+
+struct KernelTypePairHash {
+  static void HashCombine(const OpKernelType& t, std::size_t* seed) {
+    OpKernelType::Hash kernel_type_hasher;
+    (*seed) ^= kernel_type_hasher(t) + 0x9e3779b9 + (*seed << 6) + (*seed >> 2);
+  }
+
+  size_t operator()(const KernelTypePair& kernel_pair) const {
+    std::size_t seed = 0;
+    HashCombine(kernel_pair.first, &seed);
+    HashCombine(kernel_pair.second, &seed);
+    return seed;
+  }
+};
+
+template <typename InType, typename OutType>
+struct CastDataTypeFunctor {
+  HOSTDEVICE inline OutType operator()(InType in) const {
+    return static_cast<OutType>(in);
+  }
+};
+
+template <typename InType>
+struct CastDataType {
+  CastDataType(const framework::Tensor& in, framework::Tensor* out,
+               const platform::DeviceContext* ctx)
+      : in_(in), out_(out), ctx_(ctx) {}
+  const framework::Tensor in_;
+  framework::Tensor* out_;
+  const platform::DeviceContext* ctx_;
+
+  template <typename OutType>
+  void operator()() {
+    auto place = ctx_->GetPlace();
+
+    auto* in_begin = in_.data<InType>();
+    auto numel = in_.numel();
+    auto* in_end = in_begin + numel;
+    auto* out_begin = out_->mutable_data<OutType>(place);
+    if (platform::is_cpu_place(place)) {
+      platform::Transform<platform::CPUDeviceContext> trans;
+      auto* context = static_cast<const platform::CPUDeviceContext*>(ctx_);
+      trans(*context, in_begin, in_end, out_begin,
+            CastDataTypeFunctor<InType, OutType>());
+    } else {
+      // TODO(dzhwinter): enhance CopyFrom CPU<->GPU with different data type?
+      PADDLE_THROW("Unsupport CPU <-> GPU!");
+    }
+  }
+};
+
+struct CastDataLayout {
+  CastDataLayout(const framework::Tensor& in, framework::Tensor* out,
+                 const platform::DeviceContext* ctx,
+                 const std::vector<int>& axis)
+      : in_(in), out_(out), ctx_(ctx), axis_(axis) {}
+  const framework::Tensor in_;
+  framework::Tensor* out_;
+  const platform::DeviceContext* ctx_;
+  const std::vector<int> axis_;
+
+  template <typename T>
+  void operator()() {
+    auto place = ctx_->GetPlace();
+    if (platform::is_cpu_place(place)) {
+      operators::math::Transpose<platform::CPUDeviceContext, T, 4> trans4;
+      auto* context = static_cast<const platform::CPUDeviceContext*>(ctx_);
+      trans4(*context, in_, out_, axis_);
+    } else {
+      PADDLE_THROW("Unsupport CPU <-> GPU!");
+    }
+  }
+};
+
+using DataTransformMap =
+    std::unordered_map<KernelTypePair, DataTransformFn, KernelTypePairHash>;
+
+class DataTransformFnMap {
+ public:
+  static DataTransformFnMap& Instance();
+
+  bool Has(const KernelTypePair& key_pair) const {
+    return map_.find(key_pair) != map_.end();
+  }
+
+  void Insert(const OpKernelType& left, const OpKernelType& right,
+              const DataTransformFn& data_tranform_fn) {
+    Insert(std::make_pair(left, right), data_tranform_fn);
+  }
+
+  void Insert(const KernelTypePair& kernel_type_pair,
+              const DataTransformFn& data_tranform_fn) {
+    PADDLE_ENFORCE(!Has(kernel_type_pair),
+                   "KernelTypePair %s has been registered", "");
+    map_.insert({kernel_type_pair, data_tranform_fn});
+  }
+
+  const DataTransformFn& Get(const KernelTypePair& key_pair) const {
+    auto data_transformer = GetNullable(key_pair);
+    PADDLE_ENFORCE_NOT_NULL(data_transformer,
+                            "DataTransformFn should not be NULL");
+    return *data_transformer;
+  }
+
+  const DataTransformFn* GetNullable(const KernelTypePair& key_pair) const {
+    auto it = map_.find(key_pair);
+    if (it == map_.end()) {
+      return nullptr;
+    } else {
+      return &(it->second);
+    }
+  }
+
+  const DataTransformMap& Map() const { return map_; }
+
+ private:
+  DataTransformFnMap() = default;
+  DataTransformMap map_;
+  DISABLE_COPY_AND_ASSIGN(DataTransformFnMap);
+};
+
+// generate unique name with __LINE__
+// refs https://stackoverflow.com/questions/1597007
+#define TOKENPASTE(x, y) x##y
+#define TOKENPASTE2(x, y) TOKENPASTE(x, y)
+#define REGISTER_DATA_TRANSFORM_FN(from, to, fn)                              \
+  static int TOKENPASTE2(fn_, __LINE__)() {                                   \
+    ::paddle::framework::DataTransformFnMap::Instance().Insert(from, to, fn); \
+    return 0;                                                                 \
+  }                                                                           \
+  static int TOKENPASTE2(var_, __LINE__) __attribute__((unused)) =            \
+      TOKENPASTE2(fn_, __LINE__)()
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/data_transform_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 <array>
+#include <vector>
+
+#include <gtest/gtest.h>
+
+#include "paddle/framework/data_transform.h"
+#include "paddle/platform/device_context.h"
+
+namespace paddle {
+namespace framework {
+using namespace platform;
+
+/**
+ * @brief cross validation of different kernel type transform
+ *  We use four bit map represent different combination.
+ *  If the field has multiple possible value, only choose two of them.
+ *  For DataType, only test the FP32(float), FP64(double).
+ *  e.g. 0000 -> FP32, CPUPlace, kNHWC, kPlain
+ *       1111 -> FP64, GPUPlace, kNCHW, kMKLDNN
+ */
+
+std::array<proto::DataType, 2> kDataType = {
+    {proto::DataType::FP32, proto::DataType::FP64}};
+
+std::array<Place, 2> kPlace = {{CPUPlace(), CUDAPlace(0)}};
+
+std::array<DataLayout, 2> kDataLayout = {{
+    DataLayout::kNHWC, DataLayout::kNCHW,
+}};
+
+std::array<LibraryType, 2> kLibraryType = {{
+    LibraryType::kPlain, LibraryType::kMKLDNN,
+}};
+
+OpKernelType GenFromBit(const std::vector<bool> bits) {
+  return OpKernelType(kDataType[bits[0]], kPlace[bits[1]], kDataLayout[bits[2]],
+                      kLibraryType[bits[3]]);
+}
+
+int test_value = 0;
+
+auto kernel0 = GenFromBit({0, 0, 0, 0});
+auto kernel1 = GenFromBit({0, 0, 0, 1});
+auto kernel2 = GenFromBit({0, 0, 1, 0});
+auto kernel3 = GenFromBit({0, 0, 1, 1});
+
+void TransDataType_t(const platform::DeviceContext* ctx,
+                     const KernelTypePair& p, const Variable& in,
+                     Variable* out) {
+  test_value++;
+}
+
+void TransDataLayout_t(const platform::DeviceContext* ctx,
+                       const KernelTypePair& p, const Variable& in,
+                       Variable* out) {
+  test_value--;
+}
+
+void TransLibraryType_t(const platform::DeviceContext* ctx,
+                        const KernelTypePair& p, const Variable& in,
+                        Variable* out) {
+  test_value += 2;
+}
+
+}  // namespace framework
+}  // namespace paddle
+
+namespace frw = paddle::framework;
+
+REGISTER_DATA_TRANSFORM_FN(frw::kernel0, frw::kernel1, frw::TransDataType_t);
+REGISTER_DATA_TRANSFORM_FN(frw::kernel1, frw::kernel2, frw::TransDataLayout_t);
+REGISTER_DATA_TRANSFORM_FN(frw::kernel0, frw::kernel2, frw::TransLibraryType_t);
+
+TEST(DataTransform, Register) {
+  using namespace paddle::framework;
+  using namespace paddle::platform;
+
+  auto& instance = DataTransformFnMap::Instance();
+  paddle::framework::Variable in;
+  paddle::framework::Variable out;
+
+  DeviceContext* ctx = new CPUDeviceContext();
+  auto pair0 = std::make_pair(frw::kernel0, frw::kernel1);
+  instance.Get(pair0)(ctx, pair0, in, &out);
+  ASSERT_EQ(test_value, 1);
+
+  auto pair1 = std::make_pair(frw::kernel1, frw::kernel2);
+  instance.Get(pair1)(ctx, pair1, in, &out);
+  ASSERT_EQ(test_value, 0);
+
+  auto pair3 = std::make_pair(frw::kernel0, frw::kernel2);
+  instance.Get(pair3)(ctx, pair3, in, &out);
+  ASSERT_EQ(test_value, 2);
+}
+
+TEST(DataTransform, Layout) {
+  using namespace paddle::framework;
+  using namespace paddle::platform;
+
+  auto& instance = DataTransformFnMap::Instance();
+  Variable in;
+  Variable out;
+  Tensor* src = in.GetMutable<Tensor>();
+  src->mutable_data<double>(make_ddim({2, 3, 1, 2}), CPUPlace());
+  src->set_layout(DataLayout::kNHWC);
+
+  DeviceContext* ctx = new CPUDeviceContext();
+
+  {
+    auto kernel1 = GenFromBit({1, 0, 0, 0});
+    auto kernel2 = GenFromBit({1, 0, 1, 0});
+    auto pair0 = std::make_pair(kernel1, kernel2);
+    instance.Get(pair0)(ctx, pair0, in, &out);
+  }
+
+  Tensor dst = out.Get<Tensor>();
+  EXPECT_TRUE(dst.layout() != src->layout());
+}
+
+TEST(DataTransform, DataType) {
+  using namespace paddle::framework;
+  using namespace paddle::platform;
+
+  auto& instance = DataTransformFnMap::Instance();
+  DeviceContext* ctx = new CPUDeviceContext();
+
+  Variable in;
+  Variable out;
+  Tensor* src = in.GetMutable<Tensor>();
+  float* ptr = src->mutable_data<float>(make_ddim({2, 3}), CPUPlace());
+  for (int i = 0; i < 6; ++i) {
+    ptr[i] = i / 3;
+  }
+
+  {
+    auto kernel1 = GenFromBit({0, 0, 0, 0});
+    auto kernel2 = GenFromBit({1, 0, 0, 0});
+    auto pair0 = std::make_pair(kernel1, kernel2);
+    instance.Get(pair0)(ctx, pair0, in, &out);
+  }
+  Tensor dst = out.Get<Tensor>();
+  EXPECT_TRUE(dst.data<double>() != nullptr);
+}

paddle/framework/data_type.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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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 <typeindex>

paddle/framework/ddim_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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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 <sstream>
 #include <vector>
 

paddle/framework/details/op_registry.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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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
 

paddle/framework/executor.cc

@@ -14,18 +14,17 @@ limitations under the License. */
 
 #include "paddle/framework/executor.h"
 
-#include <algorithm>
-#include <iostream>
-#include <memory>
 #include <set>
-#include <vector>
 
+#include "gflags/gflags.h"
 #include "paddle/framework/feed_fetch_type.h"
 #include "paddle/framework/lod_rank_table.h"
-#include "paddle/framework/lod_tensor.h"
 #include "paddle/framework/lod_tensor_array.h"
 #include "paddle/framework/op_registry.h"
-#include "paddle/framework/scope.h"
+
+DEFINE_bool(check_nan_inf, false,
+            "Checking whether operator produce NAN/INF or not. It will be "
+            "extremely slow so please use this flag wisely.");
 
 namespace paddle {
 namespace framework {
@@ -33,13 +32,7 @@ namespace framework {
 const std::string kFeedOpType = "feed";
 const std::string kFetchOpType = "fetch";
 
-DeviceContextPool* DeviceContextPool::pool = nullptr;
-
-Executor::Executor(const std::vector<platform::Place>& places) {
-  DeviceContextPool& pool = DeviceContextPool::Get();
-  auto borrowed_contexts = pool.Borrow(places);
-  device_contexts_.swap(borrowed_contexts);
-}
+Executor::Executor(const platform::Place& place) : place_(place) {}
 
 static void CreateTensor(Variable* var, proto::VarDesc::VarType var_type) {
   if (var_type == proto::VarDesc::LOD_TENSOR) {
@@ -64,6 +57,19 @@ static void CreateTensor(Variable* var, proto::VarDesc::VarType var_type) {
   }
 }
 
+static void CheckTensorNANOrInf(const std::string& name,
+                                const framework::Tensor& tensor) {
+  if (tensor.memory_size() == 0) {
+    return;
+  }
+  if (tensor.type().hash_code() != typeid(float).hash_code() &&
+      tensor.type().hash_code() != typeid(double).hash_code()) {
+    return;
+  }
+  PADDLE_ENFORCE(!framework::HasInf(tensor), "Tensor %s has Inf", name);
+  PADDLE_ENFORCE(!framework::HasNAN(tensor), "Tensor %s has NAN", name);
+}
+
 void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
                    bool create_local_scope, bool create_vars) {
   // TODO(tonyyang-svail):
@@ -71,7 +77,6 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
   //    - will change to use multiple blocks for RNN op and Cond Op
   PADDLE_ENFORCE_LT(static_cast<size_t>(block_id), pdesc.Size());
   auto& block = pdesc.Block(block_id);
-  auto& device = device_contexts_[0];
 
   Scope* local_scope = scope;
   if (create_vars) {
@@ -107,9 +112,18 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
   for (auto& op_desc : block.AllOps()) {
     auto op = paddle::framework::OpRegistry::CreateOp(*op_desc);
     VLOG(3) << op->DebugString();
-    op->Run(*local_scope, *device);
+    op->Run(*local_scope, place_);
+    if (FLAGS_check_nan_inf) {
+      for (auto& vname : op->OutputVars(true)) {
+        auto* var = local_scope->FindVar(vname);
+        if (var == nullptr) continue;
+        if (var->IsType<framework::LoDTensor>()) {
+          CheckTensorNANOrInf(vname, var->Get<framework::LoDTensor>());
+        }
+      }
+    }
   }
-  if (create_local_scope) {
+  if (create_vars && create_local_scope) {
     scope->DeleteScope(local_scope);
   }
 }

paddle/framework/executor.h

@@ -14,9 +14,6 @@ limitations under the License. */
 
 #pragma once
 
-#include <map>
-#include <unordered_map>
-
 #include "paddle/framework/op_info.h"
 #include "paddle/framework/program_desc.h"
 #include "paddle/framework/scope.h"
@@ -26,96 +23,13 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-class DeviceContextPool {
- public:
-  static DeviceContextPool& Get() {
-    PADDLE_ENFORCE_NOT_NULL(pool, "Need to Create DeviceContextPool first!");
-    return *pool;
-  }
-
-  static DeviceContextPool& Create(const std::vector<platform::Place>& places) {
-    if (pool == nullptr) {
-      pool = new DeviceContextPool(places);
-    }
-    return *pool;
-  }
-
-  const platform::DeviceContext* Borrow(const platform::Place& place) {
-    auto range = device_contexts_.equal_range(place);
-    if (range.first == range.second) {
-      PADDLE_THROW(
-          "'Place' is not supported, Please re-compile with WITH_GPU "
-          "option");
-    }
-    return range.first->second;
-  }
-
-  std::vector<const platform::DeviceContext*> Borrow(
-      const std::vector<platform::Place>& places) {
-    PADDLE_ENFORCE_GT(places.size(), 0);
-    PADDLE_ENFORCE_LE(places.size(), device_contexts_.size());
-    std::vector<const platform::DeviceContext*> borrowed_contexts;
-    for (auto& place : places) {
-      auto range = device_contexts_.equal_range(place);
-      if (range.first == range.second) {
-        PADDLE_THROW(
-            "'Place' is not supported, Please re-compile with WITH_GPU "
-            "option");
-      }
-      // TODO(dzhwinter) : assign the first found device. Will enhanced later.
-      // device load balancer maybe useful here.
-      borrowed_contexts.emplace_back(range.first->second);
-    }
-    return borrowed_contexts;
-  }
-
-  explicit DeviceContextPool(const std::vector<platform::Place>& places) {
-    PADDLE_ENFORCE_GT(places.size(), 0);
-    for (size_t i = 0; i < places.size(); i++) {
-      if (platform::is_cpu_place(places[i])) {
-        device_contexts_.emplace(
-            places[i], new platform::CPUDeviceContext(
-                           boost::get<platform::CPUPlace>(places[i])));
-      } else if (platform::is_gpu_place(places[i])) {
-#ifdef PADDLE_WITH_CUDA
-        device_contexts_.emplace(
-            places[i], new platform::CUDADeviceContext(
-                           boost::get<platform::GPUPlace>(places[i])));
-#else
-        PADDLE_THROW(
-            "'GPUPlace' is not supported, Please re-compile with WITH_GPU "
-            "option");
-#endif
-      }
-    }
-  }
-
-  ~DeviceContextPool() {}
-
- private:
-  static DeviceContextPool* pool;
-  struct Hash {
-    std::hash<int> hash_;
-    size_t operator()(const platform::Place& place) const {
-      return hash_(place.which());
-    }
-  };
-  std::unordered_multimap<const platform::Place, const platform::DeviceContext*,
-                          Hash>
-      device_contexts_;
-  DISABLE_COPY_AND_ASSIGN(DeviceContextPool);
-};
-
 class Executor {
  public:
   // TODO(dzhwinter) : Do not rely on this function, it will be removed
   explicit Executor(const platform::DeviceContext& device)
-      : Executor(std::vector<platform::Place>({device.GetPlace()})) {}
-
-  explicit Executor(const platform::Place& place)
-      : Executor(std::vector<platform::Place>({place})) {}
+      : Executor(device.GetPlace()) {}
 
-  explicit Executor(const std::vector<platform::Place>& places);
+  explicit Executor(const platform::Place& place);
 
   /* @Brief
    * Runtime evaluation of the given ProgramDesc under certain Scope
@@ -128,7 +42,7 @@ class Executor {
            bool create_vars = true);
 
  private:
-  std::vector<const platform::DeviceContext*> device_contexts_;
+  const platform::Place place_;
 };
 
 }  // namespace framework

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

paddle/framework/grad_op_desc_maker.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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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 <string>

paddle/framework/init.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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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 <algorithm>
 #include <string>
 
-#include "paddle/framework/executor.h"
 #include "paddle/framework/init.h"
+#include "paddle/platform/device_context.h"
 #include "paddle/platform/place.h"
 #include "paddle/string/piece.h"
 
@@ -48,13 +48,13 @@ bool InitDevices(const std::vector<std::string> &devices) {
   std::vector<platform::Place> places;
   for (auto &device : devices) {
     auto p = string::Piece(device);
-    if (string::Find(p, ':', 0) == string::Piece::npos) {
+    if (string::HasPrefix(p, "CPU")) {
       places.emplace_back(platform::CPUPlace());
     } else if (string::HasPrefix(p, "GPU")) {
 #ifdef PADDLE_WITH_CUDA
       auto pos = string::RFind(p, ':', string::Piece::npos);
       auto number = device.substr(pos + 1);
-      places.emplace_back(platform::GPUPlace(std::stoi(number)));
+      places.emplace_back(platform::CUDAPlace(std::stoi(number)));
 #else
       LOG(WARNING)
           << "'GPU' is not supported, Please re-compile with WITH_GPU option";
@@ -69,10 +69,9 @@ bool InitDevices(const std::vector<std::string> &devices) {
                      return platform::is_cpu_place(place);
                    }) == places.end()) {
     places.emplace_back(platform::CPUPlace());
-    LOG(WARNING) << "Not specified any device, use CPU by Default.";
+    LOG(WARNING) << "Not specified CPU device, create CPU by Default.";
   }
-  DeviceContextPool::Create(places);
-  return true;
+  platform::DeviceContextPool::Init(places);
   return true;
 }
 

paddle/framework/init.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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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 <mutex>
 

paddle/framework/init_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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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 "paddle/framework/init.h"
@@ -23,5 +23,9 @@ TEST(Init, InitDevices) {
 #ifdef PADDLE_WITH_CUDA
   std::vector<std::string> ds2 = {"CPU", "GPU:0", "GPU:1"};
   ASSERT_EQ(InitDevices(ds2), true);
+
+  // test re-init
+  std::vector<std::string> ds3 = {"GPU:0", "GPU:1"};
+  ASSERT_EQ(InitDevices(ds3), true);
 #endif
 }

paddle/framework/library_type.h

@@ -20,7 +20,48 @@ namespace framework {
 // For more details about the design of LibraryType, Please refer to
 // https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/operator_kernel_type.md#library
 
-enum LibraryType { kPlain = 0; kMKLDNN = 1; kCUDNN = 2; }
+enum class LibraryType {
+  kPlain = 0,
+  kMKLDNN = 1,
+  kCUDNN = 2,
+};
+
+inline std::string LibraryTypeToString(const LibraryType& library_type) {
+  switch (library_type) {
+    case LibraryType::kPlain:
+      return "PLAIN";
+    case LibraryType::kMKLDNN:
+      return "MKLDNN";
+    case LibraryType::kCUDNN:
+      return "CUDNN";
+    default:
+      PADDLE_THROW("unknown LibraryType %d", static_cast<int>(library_type));
+  }
+}
+
+inline LibraryType StringToLibraryType(const char* ctype) {
+  std::string s(ctype);
+  if (s == std::string("PLAIN")) {
+    return LibraryType::kPlain;
+  } else if (s == std::string("MKLDNN")) {
+    return LibraryType::kMKLDNN;
+  } else if (s == std::string("CUDNN")) {
+    return LibraryType::kCUDNN;
+    // To be compatible with register macro.
+    // CPU, CUDA, PLAIN are same library type.
+  } else if (s == std::string("CPU")) {
+    return LibraryType::kPlain;
+  } else if (s == std::string("CUDA")) {
+    return LibraryType::kPlain;
+  } else {
+    PADDLE_THROW("Unknown LibraryType %s", s.c_str());
+  }
+}
+
+inline std::ostream& operator<<(std::ostream& out, LibraryType l) {
+  out << LibraryTypeToString(l);
+  return out;
+}
 
 }  // namespace
 }  // framework

paddle/framework/lod_rank_table.cc

-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+/* 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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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/lod_rank_table.h"
 

paddle/framework/lod_rank_table.h

-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+/* 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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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 <iosfwd>

paddle/framework/lod_tensor.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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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/lod_tensor.h"
 #include "paddle/framework/data_type.h"
@@ -189,62 +189,16 @@ void AppendLoD(LoD *lod, const LoD &lod_length) {
 
 void SerializeToStream(std::ostream &os, const LoDTensor &tensor,
                        const platform::DeviceContext &dev_ctx) {
-  // TODO(typhoonzero): serialize to ostream
-  {  // the 1st field, uint32_t version
+  {  // the 1st field, uint32_t version for LoDTensor
     constexpr uint32_t version = 0;
     os.write(reinterpret_cast<const char *>(&version), sizeof(version));
   }
-  {  // the 2nd field, tensor description
-     // int32_t  size
-     // void*    protobuf message
-    proto::TensorDesc desc;
-    desc.set_data_type(framework::ToDataType(tensor.type()));
-    auto dims = framework::vectorize(tensor.dims());
-    auto *pb_dims = desc.mutable_dims();
-    pb_dims->Resize(static_cast<int>(dims.size()), 0);
-    std::copy(dims.begin(), dims.end(), pb_dims->begin());
-    int32_t size = desc.ByteSize();
-    os.write(reinterpret_cast<const char *>(&size), sizeof(size));
-    auto out = desc.SerializeAsString();
-    os.write(out.data(), size);
-  }
-  {  // the 3rd field, tensor data
-    uint64_t size = tensor.memory_size();
-    auto *data_ptr = tensor.data<void>();
-    PADDLE_ENFORCE(size < std::numeric_limits<std::streamsize>::max(),
-                   "Index overflow when writing tensor");
-    if (platform::is_gpu_place(tensor.place())) {
-#ifdef PADDLE_WITH_CUDA
-      constexpr size_t kBufSize = 1024 * 1024 * 64;  // 64MB
-      std::unique_ptr<char[]> buf(new char[kBufSize]);
-      auto &gpu_dev_ctx =
-          static_cast<const platform::CUDADeviceContext &>(dev_ctx);
-      platform::CPUPlace cpu;
-      uintptr_t data = reinterpret_cast<uintptr_t>(data_ptr);
-      while (size != 0) {
-        size_t size_to_write = std::min(kBufSize, static_cast<size_t>(size));
-        memory::Copy(cpu, buf.get(),
-                     boost::get<platform::GPUPlace>(tensor.place()),
-                     reinterpret_cast<const void *>(data), size_to_write,
-                     gpu_dev_ctx.stream());
-        gpu_dev_ctx.Wait();
-        os.write(buf.get(), size_to_write);
-        data += size_to_write;
-        size -= size_to_write;
-      }
-#else
-      PADDLE_THROW("Unexpected branch");
-#endif
-    } else {
-      os.write(static_cast<const char *>(data_ptr),
-               static_cast<std::streamsize>(size));
-    }
-  }
-  {  // the 4th field, lod information
-     // uint64_t lod_level
-     // uint64_t lod_level_1 size in byte.
-     // int*     lod_level_1 data
-     // ...
+  {
+    // the 2st field, LoD information
+    // uint64_t lod_level
+    // uint64_t lod_level_1 size in byte.
+    // int*     lod_level_1 data
+    // ...
     auto lod = tensor.lod();
     uint64_t size = lod.size();
     os.write(reinterpret_cast<const char *>(&size), sizeof(size));
@@ -256,49 +210,19 @@ void SerializeToStream(std::ostream &os, const LoDTensor &tensor,
                static_cast<std::streamsize>(size));
     }
   }
+  // the 3st field, Tensor
+  SerializeToStream(os, static_cast<Tensor>(tensor), dev_ctx);
 }
 
 void DeserializeFromStream(std::istream &is, LoDTensor *tensor) {
-  uint32_t version;
-  is.read(reinterpret_cast<char *>(&version), sizeof(version));
-  PADDLE_ENFORCE_EQ(version, 0U, "Only version 0 is supported");
-  proto::TensorDesc desc;
-  {  // int32_t size
-     // proto buffer
-    int32_t size;
-    is.read(reinterpret_cast<char *>(&size), sizeof(size));
-    std::unique_ptr<char[]> buf(new char[size]);
-    is.read(reinterpret_cast<char *>(buf.get()), size);
-    PADDLE_ENFORCE(desc.ParseFromArray(buf.get(), size),
-                   "Cannot parse tensor desc");
-  }
-  {  // read tensor
-    std::vector<int64_t> dims;
-    dims.reserve(static_cast<size_t>(desc.dims().size()));
-    std::copy(desc.dims().begin(), desc.dims().end(), std::back_inserter(dims));
-    tensor->Resize(framework::make_ddim(dims));
-
-    void *buf;
-    platform::Place cpu = platform::CPUPlace();
-    switch (desc.data_type()) {
-      case proto::FP32:
-        buf = tensor->mutable_data<float>(cpu);
-        break;
-      case proto::FP64:
-        buf = tensor->mutable_data<double>(cpu);
-        break;
-      case proto::INT32:
-        buf = tensor->mutable_data<int>(cpu);
-        break;
-      case proto::INT64:
-        buf = tensor->mutable_data<int64_t>(cpu);
-        break;
-      default:
-        PADDLE_THROW("DataType %d not supported", desc.data_type());
-    }
-    is.read(static_cast<char *>(buf), tensor->memory_size());
-  }
-  {  // read lod
+  {
+    // the 1st field, unit32_t version for SelectedRows
+    uint32_t version;
+    is.read(reinterpret_cast<char *>(&version), sizeof(version));
+    PADDLE_ENFORCE_EQ(version, 0U, "Only version 0 is supported");
+  }
+  {
+    // the 2st field, LoD information
     uint64_t lod_level;
     is.read(reinterpret_cast<char *>(&lod_level), sizeof(lod_level));
     auto &lod = *tensor->mutable_lod();
@@ -312,6 +236,8 @@ void DeserializeFromStream(std::istream &is, LoDTensor *tensor) {
       lod[i] = tmp;
     }
   }
+  // the 3st filed, Tensor
+  DeserializeFromStream(is, static_cast<Tensor *>(tensor));
 }
 
 }  // namespace framework

paddle/framework/lod_tensor.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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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
 

paddle/framework/lod_tensor_array.h

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

paddle/framework/lod_tensor_test.cc

@@ -126,6 +126,20 @@ TEST_F(LoDTensorTester, ShrinkInLevel) {
   EXPECT_NE(t1.data<float>(), lod_tensor_.data<float>());
 }
 
+TEST_F(LoDTensorTester, SerializeAndDeserialize) {
+  LoDTensor dst_tensor;
+  platform::CPUDeviceContext cpu_ctx((platform::CPUPlace()));
+  std::ostringstream oss;
+  SerializeToStream(oss, lod_tensor_, cpu_ctx);
+  std::istringstream iss(oss.str());
+  DeserializeFromStream(iss, &dst_tensor);
+  float* dst_ptr = dst_tensor.mutable_data<float>(platform::CPUPlace());
+  for (int i = 0; i < kLodTensorSize; ++i) {
+    EXPECT_EQ(dst_ptr[i], i);
+  }
+  EXPECT_EQ(dst_tensor.lod(), lod_tensor_.lod());
+}
+
 TEST(LodExpand, test) {
   LoD lod{{0, 2}};
   LoDTensor tensor;

paddle/framework/lod_tensor_test.cu

@@ -27,7 +27,7 @@ __global__ void test(size_t* a, int size) {
 
 TEST(LoDTensor, LoDInGPU) {
   paddle::framework::LoDTensor lod_tensor;
-  paddle::platform::GPUPlace place(0);
+  paddle::platform::CUDAPlace place(0);
 
   paddle::framework::LoD src_lod;
   src_lod.push_back(std::vector<size_t>{0, 2, 4, 6, 8, 10, 12, 14});

paddle/framework/op_desc.cc

@@ -88,6 +88,14 @@ OpDesc::OpDesc(const std::string &type, const VariableNameMap &inputs,
   need_update_ = true;
 }
 
+void OpDesc::CopyFrom(const OpDesc &op_desc) {
+  desc_.set_type(op_desc.Type());
+  inputs_ = op_desc.inputs_;
+  outputs_ = op_desc.outputs_;
+  attrs_ = op_desc.attrs_;
+  need_update_ = true;
+}
+
 OpDesc::OpDesc(const proto::OpDesc &desc, ProgramDesc *prog)
     : desc_(desc), need_update_(false) {
   // restore inputs_

paddle/framework/op_desc.h

@@ -35,6 +35,8 @@ class OpDesc {
 
   OpDesc(const proto::OpDesc &desc, ProgramDesc *prog);
 
+  void CopyFrom(const OpDesc &op_desc);
+
   proto::OpDesc *Proto();
 
   std::string Type() const { return desc_.type(); }

paddle/framework/op_info.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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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_info.h"
 

paddle/framework/op_info.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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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 <functional>

paddle/framework/op_kernel_type.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/data_layout.h"
+#include "paddle/framework/data_type.h"
+#include "paddle/framework/library_type.h"
+#include "paddle/platform/device_context.h"
+#include "paddle/platform/place.h"
+
+namespace paddle {
+namespace framework {
+
+struct OpKernelType {
+  struct Hash {
+    size_t operator()(const OpKernelType& key) const {
+      int place = key.place_.which() + (1 << LEFT_SHIFT);
+      int data_type =
+          static_cast<int>(key.data_type_) + (1 << (LEFT_SHIFT + 1));
+      int data_layout =
+          static_cast<int>(key.data_layout_) + (1 << (LEFT_SHIFT + 2));
+      int library_type =
+          static_cast<int>(key.library_type_) + (1 << (LEFT_SHIFT + 3));
+      std::hash<int> hasher;
+      return hasher(place + data_type + data_layout + library_type);
+    }
+  };
+
+  // place, data_type, library_type kinds less than 2^8
+  constexpr static int LEFT_SHIFT = 8;
+
+  proto::DataType data_type_;
+  DataLayout data_layout_;
+  platform::Place place_;
+  LibraryType library_type_;
+
+  OpKernelType(proto::DataType data_type, platform::Place place,
+               DataLayout data_layout = DataLayout::kAnyLayout,
+               LibraryType library_type = LibraryType::kPlain)
+      : data_type_(data_type),
+        data_layout_(data_layout),
+        place_(place),
+        library_type_(library_type) {}
+
+  OpKernelType(proto::DataType data_type,
+               const platform::DeviceContext& dev_ctx,
+               DataLayout data_layout = DataLayout::kAnyLayout,
+               LibraryType library_type = LibraryType::kPlain)
+      : data_type_(data_type),
+        data_layout_(data_layout),
+        place_(dev_ctx.GetPlace()),
+        library_type_(library_type) {}
+
+  bool operator==(const OpKernelType& o) const {
+    return platform::places_are_same_class(place_, o.place_) &&
+           data_type_ == o.data_type_ && data_layout_ == o.data_layout_ &&
+           library_type_ == o.library_type_;
+  }
+
+  bool operator!=(const OpKernelType& o) const { return !(*this == o); }
+};
+
+inline std::ostream& operator<<(std::ostream& os,
+                                const OpKernelType& kernel_key) {
+  os << "data_type[" << kernel_key.data_type_ << "]:data_layout["
+     << kernel_key.data_layout_ << "]:place[" << kernel_key.place_
+     << "]:library_type[" << kernel_key.library_type_ << "]";
+  return os;
+}
+
+inline std::string KernelTypeToString(const OpKernelType& kernel_key) {
+  std::ostringstream stream;
+  stream << kernel_key;
+  return stream.str();
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/op_kernel_type_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/framework/op_kernel_type.h"
+#include <gtest/gtest.h>
+#include <iostream>
+
+TEST(OpKernelType, ToString) {
+  using OpKernelType = paddle::framework::OpKernelType;
+  using DataType = paddle::framework::proto::DataType;
+  using CPUPlace = paddle::platform::CPUPlace;
+  using DataLayout = paddle::framework::DataLayout;
+  using LibraryType = paddle::framework::LibraryType;
+
+  OpKernelType op_kernel_type(DataType::FP32, CPUPlace(), DataLayout::kNCHW,
+                              LibraryType::kCUDNN);
+
+  ASSERT_EQ(
+      paddle::framework::KernelTypeToString(op_kernel_type),
+      "data_type[5]:data_layout[NCHW]:place[CPUPlace]:library_type[CUDNN]");
+}
+
+TEST(OpKernelType, Hash) {
+  using OpKernelType = paddle::framework::OpKernelType;
+  using DataType = paddle::framework::proto::DataType;
+  using CPUPlace = paddle::platform::CPUPlace;
+  using CUDAPlace = paddle::platform::CUDAPlace;
+  using DataLayout = paddle::framework::DataLayout;
+  using LibraryType = paddle::framework::LibraryType;
+
+  OpKernelType op_kernel_type_1(DataType::FP32, CPUPlace(), DataLayout::kNCHW,
+                                LibraryType::kCUDNN);
+  OpKernelType op_kernel_type_2(DataType::FP32, CUDAPlace(0), DataLayout::kNCHW,
+                                LibraryType::kCUDNN);
+
+  OpKernelType::Hash hasher;
+  ASSERT_NE(hasher(op_kernel_type_1), hasher(op_kernel_type_2));
+}

paddle/framework/op_registry.h

@@ -61,17 +61,6 @@ struct OperatorRegistrar : public Registrar {
 
 class OpRegistry {
  public:
-  template <typename OpType, typename ProtoMakerType, typename GradOpType>
-  static void RegisterOp(const std::string& op_type,
-                         const std::string& grad_op_type) {
-    OperatorRegistrar<OpType, ProtoMakerType> reg(op_type.c_str());
-    reg.info.grad_op_type_ = grad_op_type;
-    // register gradient op
-    if (!grad_op_type.empty()) {
-      OperatorRegistrar<GradOpType> grad_reg(grad_op_type.c_str());
-    }
-  }
-
   static std::unique_ptr<OperatorBase> CreateOp(const std::string& type,
                                                 const VariableNameMap& inputs,
                                                 const VariableNameMap& outputs,
@@ -90,30 +79,31 @@ struct OpKernelRegistrarFunctor<PlaceType, false, I, KernelTypes...> {
   using KERNEL_TYPE =
       typename std::tuple_element<I, std::tuple<KernelTypes...>>::type;
 
-  void operator()(const char* op_type) const {
+  void operator()(const char* op_type, const char* library_type) const {
     using T = typename KERNEL_TYPE::ELEMENT_TYPE;
-    OpKernelType key(ToDataType(std::type_index(typeid(T))), PlaceType());
+    OpKernelType key(ToDataType(std::type_index(typeid(T))), PlaceType(),
+                     DataLayout::kAnyLayout, StringToLibraryType(library_type));
     OperatorWithKernel::AllOpKernels()[op_type][key].reset(new KERNEL_TYPE);
 
     constexpr auto size = std::tuple_size<std::tuple<KernelTypes...>>::value;
     OpKernelRegistrarFunctor<PlaceType, I + 1 == size, I + 1, KernelTypes...>
         func;
-    func(op_type);
+    func(op_type, library_type);
   }
 };
 
 template <typename PlaceType, size_t I, typename... KernelType>
 struct OpKernelRegistrarFunctor<PlaceType, true, I, KernelType...> {
-  void operator()(const char* op_type) const {}
+  void operator()(const char* op_type, const char* library_type) const {}
 };
 
 // User can register many kernel in one place. The data type could be different.
 template <typename PlaceType, typename... KernelType>
 class OpKernelRegistrar : public Registrar {
  public:
-  explicit OpKernelRegistrar(const char* op_type) {
+  explicit OpKernelRegistrar(const char* op_type, const char* library_type) {
     OpKernelRegistrarFunctor<PlaceType, false, 0, KernelType...> func;
-    func(op_type);
+    func(op_type, library_type);
   }
 };
 
@@ -192,14 +182,15 @@ class OpKernelRegistrar : public Registrar {
       __reg_op_kernel_##op_type##_##DEVICE_TYPE##__,                      \
       "REGISTER_OP_KERNEL must be called in global namespace");           \
   static ::paddle::framework::OpKernelRegistrar<place_class, __VA_ARGS__> \
-      __op_kernel_registrar_##op_type##_##DEVICE_TYPE##__(#op_type);      \
+      __op_kernel_registrar_##op_type##_##DEVICE_TYPE##__(#op_type,       \
+                                                          #DEVICE_TYPE);  \
   int TouchOpKernelRegistrar_##op_type##_##DEVICE_TYPE() {                \
     __op_kernel_registrar_##op_type##_##DEVICE_TYPE##__.Touch();          \
     return 0;                                                             \
   }
 
 #define REGISTER_OP_CUDA_KERNEL(op_type, ...) \
-  REGISTER_OP_KERNEL(op_type, CUDA, ::paddle::platform::GPUPlace, __VA_ARGS__)
+  REGISTER_OP_KERNEL(op_type, CUDA, ::paddle::platform::CUDAPlace, __VA_ARGS__)
 
 #define REGISTER_OP_CPU_KERNEL(op_type, ...) \
   REGISTER_OP_KERNEL(op_type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__)

paddle/framework/op_registry_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/framework/op_registry.h"
 #include <gtest/gtest.h>
 
@@ -8,8 +22,7 @@ namespace framework {
 class CosineOp : public OperatorBase {
  public:
   using OperatorBase::OperatorBase;
-  void Run(const Scope& scope,
-           const platform::DeviceContext& dev_ctx) const override {}
+  void Run(const Scope& scope, const platform::Place& place) const override {}
 };
 
 class CosineOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
@@ -28,8 +41,7 @@ class CosineOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
 class MyTestOp : public OperatorBase {
  public:
   using OperatorBase::OperatorBase;
-  void Run(const Scope& scope,
-           const platform::DeviceContext& dev_ctx) const override {}
+  void Run(const Scope& scope, const platform::Place& place) const override {}
 };
 
 class MyTestOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
@@ -76,8 +88,8 @@ TEST(OpRegistry, CreateOp) {
 
   auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
   paddle::framework::Scope scope;
-  paddle::platform::CPUDeviceContext dev_ctx;
-  op->Run(scope, dev_ctx);
+  paddle::platform::CPUPlace cpu_place;
+  op->Run(scope, cpu_place);
   float scale_get = op->Attr<float>("scale");
   ASSERT_EQ(scale_get, scale);
 }
@@ -117,8 +129,8 @@ TEST(OpRegistry, DefaultValue) {
 
   auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
   paddle::framework::Scope scope;
-  paddle::platform::CPUDeviceContext dev_ctx;
-  op->Run(scope, dev_ctx);
+  paddle::platform::CPUPlace cpu_place;
+  op->Run(scope, cpu_place);
   ASSERT_EQ(op->Attr<float>("scale"), 1.0);
 }
 
@@ -167,9 +179,9 @@ TEST(OpRegistry, CustomChecker) {
   attr->set_type(paddle::framework::proto::AttrType::INT);
   attr->set_i(4);
   auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
-  paddle::platform::CPUDeviceContext dev_ctx;
+  paddle::platform::CPUPlace cpu_place;
   paddle::framework::Scope scope;
-  op->Run(scope, dev_ctx);
+  op->Run(scope, cpu_place);
   int test_attr = op->Attr<int>("test_attr");
   ASSERT_EQ(test_attr, 4);
 }
@@ -184,3 +196,71 @@ TEST(OperatorRegistrar, Test) {
   using namespace paddle::framework;
   OperatorRegistrar<CosineOpComplete, CosineOpProtoAndCheckerMaker> reg("cos");
 }
+
+namespace paddle {
+namespace framework {
+
+class OpKernelTestMaker : public OpProtoAndCheckerMaker {
+ public:
+  OpKernelTestMaker(OpProto* proto, OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddComment("NoGradOp, same input output. no Grad");
+  }
+};
+
+class OpWithKernelTest : public OperatorWithKernel {
+ public:
+  using OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(InferShapeContext* ctx) const override {}
+
+  framework::OpKernelType GetActualKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(proto::DataType::FP32, ctx.device_context());
+  }
+};
+
+template <typename DeviceContext, typename T>
+class OpKernelTest : public paddle::framework::OpKernel<T> {
+ public:
+  void Compute(const paddle::framework::ExecutionContext& ctx) const {}
+};
+
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_OP_WITHOUT_GRADIENT(op_with_kernel,
+                             paddle::framework::OpWithKernelTest,
+                             paddle::framework::OpKernelTestMaker);
+REGISTER_OP_CPU_KERNEL(
+    op_with_kernel,
+    paddle::framework::OpKernelTest<paddle::platform::CPUDeviceContext, float>);
+
+REGISTER_OP_CUDA_KERNEL(op_with_kernel,
+                        paddle::framework::OpKernelTest<
+                            paddle::platform::CUDADeviceContext, float>);
+
+TEST(OperatorRegistrar, CPU) {
+  paddle::framework::proto::OpDesc op_desc;
+  paddle::platform::CPUPlace cpu_place;
+  paddle::framework::Scope scope;
+
+  op_desc.set_type("op_with_kernel");
+  auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
+
+  op->Run(scope, cpu_place);
+}
+
+#ifdef PADDLE_WITH_CUDA
+TEST(OperatorRegistrar, CUDA) {
+  paddle::framework::proto::OpDesc op_desc;
+  paddle::platform::CUDAPlace cuda_place(0);
+  paddle::framework::Scope scope;
+
+  op_desc.set_type("op_with_kernel");
+  auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
+
+  op->Run(scope, cuda_place);
+}
+#endif

paddle/framework/operator.cc

@@ -12,10 +12,13 @@ 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/operator.h"
 #include <algorithm>
 #include <atomic>
+
+#include "paddle/framework/data_transform.h"
+#include "paddle/framework/executor.h"
 #include "paddle/framework/lod_tensor_array.h"
+#include "paddle/framework/operator.h"
 #include "paddle/framework/shape_inference.h"
 #include "paddle/framework/var_type.h"
 
@@ -240,12 +243,6 @@ std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
   return res;
 }
 
-std::ostream& operator<<(std::ostream& os, const OpKernelType& kernel_key) {
-  os << "place[" << kernel_key.place_ << "]:data_type[" << kernel_key.data_type_
-     << "]";
-  return os;
-}
-
 bool OpSupportGPU(const std::string& op_type) {
   auto& all_kernels = OperatorWithKernel::AllOpKernels();
   auto it = all_kernels.find(op_type);
@@ -387,12 +384,30 @@ class RuntimeInferShapeContext : public InferShapeContext {
   const Scope& scope_;
 };
 
+const platform::DeviceContext* GetDeviceContext(
+    framework::KernelTypePair& kernel_pair) {
+  auto& actual_kernel_key = kernel_pair.first;
+  auto& expected_kernel_key = kernel_pair.second;
+  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+
+  if (platform::is_gpu_place(actual_kernel_key.place_) &&
+      platform::is_cpu_place(expected_kernel_key.place_)) {
+    return pool.Get(actual_kernel_key.place_);
+  } else if (platform::is_cpu_place(actual_kernel_key.place_) &&
+             platform::is_gpu_place(expected_kernel_key.place_)) {
+    return pool.Get(expected_kernel_key.place_);
+  } else {
+    PADDLE_THROW(
+        "Currently, model parallelism is only supported between CPU and CUDA");
+  }
+}
+
 void OperatorWithKernel::Run(const Scope& scope,
-                             const platform::DeviceContext& dev_ctx) const {
+                             const platform::Place& place) const {
   RuntimeInferShapeContext infer_shape_ctx(*this, scope);
   this->InferShape(&infer_shape_ctx);
-
-  ExecutionContext ctx(*this, scope, dev_ctx);
+  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+  auto dev_ctx = pool.Get(place);
 
   // check if op[type] has kernel registered.
   auto& all_op_kernels = AllOpKernels();
@@ -404,19 +419,71 @@ void OperatorWithKernel::Run(const Scope& scope,
 
   // check if op[type] have kernel for kernel_key
   OpKernelMap& kernels = kernels_iter->second;
-  auto kernel_key = GetKernelType(ctx);
-  auto kernel_iter = kernels.find(kernel_key);
+
+  ExecutionContext ctx(*this, scope, *dev_ctx);
+  auto actual_kernel_key = GetActualKernelType(ctx);
+  auto expected_kernel_key = GetExpectedKernelType(actual_kernel_key);
+  auto kernel_iter = kernels.find(expected_kernel_key);
 
   if (kernel_iter == kernels.end()) {
-    PADDLE_THROW("The operator %s does not support %s", type_, kernel_key);
+    PADDLE_THROW("The operator %s does not support %s", type_,
+                 expected_kernel_key);
+  }
+
+  if (actual_kernel_key == expected_kernel_key) {
+    PADDLE_ENFORCE_EQ(actual_kernel_key.place_, expected_kernel_key.place_,
+                      "Currently, model parallelism is only supported between "
+                      "CPU and other devices. For example, multi-GPU model "
+                      "parallelism will failed.");
+  } else {
+    auto kernel_pair = std::make_pair(actual_kernel_key, expected_kernel_key);
+    const DataTransformFn* trans_fun =
+        DataTransformFnMap::Instance().GetNullable(kernel_pair);
+    if (trans_fun) {
+      auto input_vars = this->InputVars();
+      // TODO(qijun) filter the input vars that do not need to be transformed
+
+      // filter vars that has been transformed
+      std::vector<std::string> need_trans;
+      for (auto var_name : input_vars) {
+        auto var_name_trans =
+            var_name + framework::KernelTypeToString(expected_kernel_key);
+        if (!scope.FindVar(var_name_trans)) {
+          const_cast<Scope&>(scope).Var(var_name_trans);
+          need_trans.push_back(var_name);
+        }
+      }
+
+      if (!need_trans.empty()) {
+        auto trans_dev_ctx = GetDeviceContext(kernel_pair);
+
+        // Wait for transform starting
+        dev_ctx->Wait();
+
+        for (auto var_name : need_trans) {
+          (*trans_fun)(trans_dev_ctx, kernel_pair, *(scope.FindVar(var_name)),
+                       scope.FindVar(var_name + framework::KernelTypeToString(
+                                                    expected_kernel_key)));
+        }
+        // Wait for data transform finishing
+        trans_dev_ctx->Wait();
+      }
+    }
   }
 
   kernel_iter->second->Compute(ctx);
 }
-OpKernelType OperatorWithKernel::GetKernelType(
+
+OpKernelType OperatorWithKernel::GetActualKernelType(
     const ExecutionContext& ctx) const {
   return OpKernelType(IndicateDataType(ctx), ctx.GetPlace());
 }
+
+OpKernelType OperatorWithKernel::GetExpectedKernelType(
+    const OpKernelType& actual_kernel_type) const {
+  return actual_kernel_type;
+}
+
 proto::DataType OperatorWithKernel::IndicateDataType(
     const ExecutionContext& ctx) const {
   auto& scope = ctx.scope();

paddle/framework/operator.h

@@ -23,15 +23,14 @@ limitations under the License. */
 #include "glog/logging.h"  // For VLOG
 #include "paddle/framework/attribute.h"
 #include "paddle/framework/block_desc.h"
-#include "paddle/framework/data_type.h"
 #include "paddle/framework/framework.pb.h"
 #include "paddle/framework/lod_tensor.h"
 #include "paddle/framework/op_info.h"
+#include "paddle/framework/op_kernel_type.h"
 #include "paddle/framework/scope.h"
 #include "paddle/framework/selected_rows.h"
 #include "paddle/framework/tensor.h"
 #include "paddle/platform/device_context.h"
-#include "paddle/platform/place.h"
 #include "paddle/platform/variant.h"
 #include "paddle/utils/Error.h"
 
@@ -53,6 +52,11 @@ constexpr char kGradVarSuffix[] = "@GRAD";
 /// Variables with this suffix are supposed to be filled up with zeros.
 constexpr char kZeroVarSuffix[] = "@ZERO";
 
+// define some kernel hint
+const std::string kUseCPU = "use_cpu";
+const std::string kUseCUDNN = "use_cudnn";
+const std::string kUseMKLDNN = "use_mkldnn";
+
 inline std::string GradVarName(const std::string& var_name) {
   return var_name + kGradVarSuffix;
 }
@@ -83,8 +87,10 @@ class OperatorBase {
   virtual std::string DebugString() const;
 
   /// Net will call this function to Run an op.
-  virtual void Run(const Scope& scope,
-                   const platform::DeviceContext& dev_ctx) const = 0;
+  virtual void Run(const Scope& scope, const platform::Place& place) const = 0;
+
+  // FIXME(typhoonzero): this is only used for recv_op to stop event_loop.
+  virtual void Stop() {}
 
   virtual bool IsNetOp() const { return false; }
 
@@ -159,8 +165,7 @@ class OperatorBase {
 class NOP : public OperatorBase {
  public:
   using OperatorBase::OperatorBase;
-  void Run(const Scope& scope,
-           const platform::DeviceContext& dev_ctx) const override {}
+  void Run(const Scope& scope, const platform::Place& place) const override {}
   std::unique_ptr<OperatorBase> Clone() const override {
     return std::unique_ptr<OperatorBase>(new NOP(*this));
   }
@@ -345,34 +350,6 @@ class OpKernel : public OpKernelBase {
   using ELEMENT_TYPE = T;
 };
 
-struct OpKernelType {
-  struct Hash {
-    std::hash<int> hash_;
-    size_t operator()(const OpKernelType& key) const {
-      int place = key.place_.which();
-      int data_type = static_cast<int>(key.data_type_);
-      int pre_hash = data_type << NUM_PLACE_TYPE_LIMIT_IN_BIT |
-                     (place & ((1 << NUM_PLACE_TYPE_LIMIT_IN_BIT) - 1));
-      return hash_(pre_hash);
-    }
-  };
-
-  platform::Place place_;
-  proto::DataType data_type_;
-
-  OpKernelType(proto::DataType data_type, platform::Place place)
-      : place_(place), data_type_(data_type) {}
-
-  OpKernelType(proto::DataType data_type,
-               const platform::DeviceContext& dev_ctx)
-      : place_(dev_ctx.GetPlace()), data_type_(data_type) {}
-
-  bool operator==(const OpKernelType& o) const {
-    return platform::places_are_same_class(place_, o.place_) &&
-           data_type_ == o.data_type_;
-  }
-};
-
 class OperatorWithKernel : public OperatorBase {
  public:
   using OpKernelMap =
@@ -383,8 +360,7 @@ class OperatorWithKernel : public OperatorBase {
                      const VariableNameMap& outputs, const AttributeMap& attrs)
       : OperatorBase(type, inputs, outputs, attrs) {}
 
-  void Run(const Scope& scope,
-           const platform::DeviceContext& dev_ctx) const final;
+  void Run(const Scope& scope, const platform::Place& place) const final;
 
   static std::unordered_map<std::string /* op_type */, OpKernelMap>&
   AllOpKernels() {
@@ -405,7 +381,9 @@ class OperatorWithKernel : public OperatorBase {
   }
 
  protected:
-  virtual OpKernelType GetKernelType(const ExecutionContext& ctx) const;
+  virtual OpKernelType GetActualKernelType(const ExecutionContext& ctx) const;
+  virtual OpKernelType GetExpectedKernelType(
+      const OpKernelType& actual_kernel_type) const;
 
  private:
   // indicate kernel DataType by input data. Defaultly all input data must be
@@ -413,8 +391,6 @@ class OperatorWithKernel : public OperatorBase {
   proto::DataType IndicateDataType(const ExecutionContext& ctx) const;
 };
 
-std::ostream& operator<<(std::ostream& os, const OpKernelType& kernel_key);
-
 extern bool OpSupportGPU(const std::string& op_type);
 
 }  // namespace framework

paddle/framework/operator_test.cc

@@ -11,11 +11,12 @@ 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/operator.h"
 #include "gtest/gtest.h"
+
+#include "paddle/framework/init.h"
 #include "paddle/framework/op_info.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/framework/operator.h"
 
 namespace paddle {
 namespace framework {
@@ -27,8 +28,7 @@ class OpWithoutKernelTest : public OperatorBase {
   OpWithoutKernelTest(const std::string& type, const VariableNameMap& inputs,
                       const VariableNameMap& outputs, const AttributeMap& attrs)
       : OperatorBase(type, inputs, outputs, attrs), x(1) {}
-  void Run(const Scope& scope,
-           const platform::DeviceContext& dev_ctx) const override {
+  void Run(const Scope& scope, const platform::Place& place) const override {
     ++op_run_num;
     ASSERT_EQ(static_cast<int>(inputs_.size()), 1);
     ASSERT_EQ(static_cast<int>(outputs_.size()), 1);
@@ -41,10 +41,9 @@ class OpWithoutKernelTest : public OperatorBase {
   int x{0};
 };
 
-class OpeWithoutKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
+class OpWithoutKernelCheckerMaker : public OpProtoAndCheckerMaker {
  public:
-  OpeWithoutKernelTestProtoAndCheckerMaker(OpProto* proto,
-                                           OpAttrChecker* op_checker)
+  OpWithoutKernelCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("input", "input of test op");
     AddOutput("output", "output of test op");
@@ -65,11 +64,12 @@ static void BuildVar(const std::string& param_name,
   }
 }
 
-REGISTER_OP_WITHOUT_GRADIENT(
-    test_operator, paddle::framework::OpWithoutKernelTest,
-    paddle::framework::OpeWithoutKernelTestProtoAndCheckerMaker);
+REGISTER_OP_WITHOUT_GRADIENT(test_operator,
+                             paddle::framework::OpWithoutKernelTest,
+                             paddle::framework::OpWithoutKernelCheckerMaker);
 
 TEST(OperatorBase, all) {
+  paddle::framework::InitDevices({"CPU"});
   paddle::framework::proto::OpDesc op_desc;
   op_desc.set_type("test_operator");
   BuildVar("input", {"IN1"}, op_desc.add_inputs());
@@ -80,13 +80,13 @@ TEST(OperatorBase, all) {
   attr->set_type(paddle::framework::proto::AttrType::FLOAT);
   attr->set_f(3.14);
 
-  paddle::platform::CPUDeviceContext device_context;
+  paddle::platform::CPUPlace cpu_place;
   paddle::framework::Scope scope;
 
   auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
   scope.Var("OUT1");
   ASSERT_EQ(paddle::framework::op_run_num, 0);
-  op->Run(scope, device_context);
+  op->Run(scope, cpu_place);
   ASSERT_EQ(paddle::framework::op_run_num, 1);
 }
 
@@ -114,7 +114,7 @@ class OpWithKernelTest : public OperatorWithKernel {
 
  protected:
   void InferShape(framework::InferShapeContext* ctx) const override {}
-  OpKernelType GetKernelType(const ExecutionContext& ctx) const override {
+  OpKernelType GetActualKernelType(const ExecutionContext& ctx) const override {
     return OpKernelType(proto::DataType::FP32, ctx.GetPlace());
   }
 };
@@ -123,7 +123,6 @@ template <typename T1, typename T2>
 class CPUKernelTest : public OpKernel<float> {
  public:
   void Compute(const ExecutionContext& ctx) const {
-    std::cout << "this is cpu kernel" << std::endl;
     std::cout << ctx.op().DebugString() << std::endl;
     cpu_kernel_run_num++;
     ASSERT_EQ(ctx.op().Input("x"), "IN1");
@@ -195,6 +194,7 @@ REGISTER_OP_CPU_KERNEL(op_with_kernel,
 
 // test with single input
 TEST(OpKernel, all) {
+  paddle::framework::InitDevices({"CPU"});
   paddle::framework::proto::OpDesc op_desc;
   op_desc.set_type("op_with_kernel");
   BuildVar("x", {"IN1"}, op_desc.add_inputs());
@@ -205,12 +205,12 @@ TEST(OpKernel, all) {
   attr->set_type(paddle::framework::proto::AttrType::FLOAT);
   attr->set_f(3.14);
 
-  paddle::platform::CPUDeviceContext cpu_device_context;
+  paddle::platform::CPUPlace cpu_place;
   paddle::framework::Scope scope;
 
   auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
   ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 0);
-  op->Run(scope, cpu_device_context);
+  op->Run(scope, cpu_place);
   ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 1);
 }
 
@@ -224,7 +224,9 @@ REGISTER_OP_CPU_KERNEL(op_multi_inputs_with_kernel,
 TEST(OpKernel, multi_inputs) {
   using namespace paddle::framework;
 
+  paddle::framework::InitDevices({"CPU"});
   proto::OpDesc op_desc;
+
   op_desc.set_type("op_multi_inputs_with_kernel");
   BuildVar("xs", {"x0", "x1", "x2"}, op_desc.add_inputs());
   BuildVar("k", {"k0"}, op_desc.add_inputs());
@@ -235,7 +237,7 @@ TEST(OpKernel, multi_inputs) {
   attr->set_type(paddle::framework::proto::AttrType::FLOAT);
   attr->set_f(3.14);
 
-  paddle::platform::CPUDeviceContext cpu_device_context;
+  paddle::platform::CPUPlace cpu_place;
   paddle::framework::Scope scope;
   scope.Var("x0")->GetMutable<LoDTensor>();
   scope.Var("x1")->GetMutable<LoDTensor>();
@@ -245,7 +247,7 @@ TEST(OpKernel, multi_inputs) {
   scope.Var("y1")->GetMutable<LoDTensor>();
 
   auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
-  op->Run(scope, cpu_device_context);
+  op->Run(scope, cpu_place);
 }
 
 class OperatorClone : public paddle::framework::OperatorBase {
@@ -257,10 +259,11 @@ class OperatorClone : public paddle::framework::OperatorBase {
                 const paddle::framework::AttributeMap& attrs)
       : OperatorBase(type, inputs, outputs, attrs) {}
   void Run(const paddle::framework::Scope& scope,
-           const paddle::platform::DeviceContext& dev_ctx) const override {}
+           const paddle::platform::Place& place) const override {}
 };
 
 TEST(Operator, Clone) {
+  paddle::framework::InitDevices({"CPU"});
   OperatorClone a("ABC", paddle::framework::VariableNameMap{},
                   paddle::framework::VariableNameMap{},
                   paddle::framework::AttributeMap{});

paddle/framework/program_desc_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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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/program_desc.h"
 #include "gtest/gtest.h"

paddle/framework/prune_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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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/prune.h"
 

paddle/framework/scope.cc

@@ -74,17 +74,9 @@ void Scope::DropKids() {
   kids_.clear();
 }
 
-std::vector<std::string> Scope::GetAllNames(bool recursive) const {
-  std::vector<std::string> known_vars(vars_.size());
-
-  if (recursive) {
-    for (auto& kid : kids_) {
-      auto kid_vars = kid->GetAllNames();
-      for (auto& p : kid_vars) {
-        known_vars.emplace_back(p);
-      }
-    }
-  }
+std::vector<std::string> Scope::LocalVarNames() const {
+  std::vector<std::string> known_vars;
+  known_vars.reserve(this->vars_.size());
   for (auto& p : vars_) {
     known_vars.emplace_back(p.first);
   }

paddle/framework/scope.h

@@ -66,7 +66,7 @@ class Scope {
   void DropKids();
 
   // enumerate all the variables current contains.
-  std::vector<std::string> GetAllNames(bool recursive = false) const;
+  std::vector<std::string> LocalVarNames() const;
 
   // Rename variable to a new name
   void Rename(const std::string& origin_name,

paddle/framework/scope_test.cc

@@ -61,7 +61,7 @@ TEST(Scope, GetAllNames) {
   Variable* v = s.Var("a");
   EXPECT_EQ(&s, s.FindScope(v));
 
-  std::vector<std::string> ans = s.GetAllNames();
+  std::vector<std::string> ans = s.LocalVarNames();
   std::string str;
   for (auto& var : ans) {
     str += var;

paddle/framework/selected_rows.cc

@@ -12,5 +12,58 @@ limitations under the License. */
 #include "paddle/framework/selected_rows.h"
 
 namespace paddle {
-namespace framework {}  // namespace framework
+namespace framework {
+void SerializeToStream(std::ostream& os, const SelectedRows& selected_rows,
+                       const platform::DeviceContext& dev_ctx) {
+  {  // the 1st field, uint32_t version
+    constexpr uint32_t version = 0;
+    os.write(reinterpret_cast<const char*>(&version), sizeof(version));
+  }
+  {
+    // the 2st field, rows information
+    auto& rows = selected_rows.rows();
+    uint64_t size = rows.size();
+    os.write(reinterpret_cast<const char*>(&size), sizeof(size));
+    for (uint64_t i = 0; i < size; ++i) {
+      os.write(reinterpret_cast<const char*>(&rows[i]), sizeof(rows[i]));
+    }
+  }
+  {
+    // the 3st field, the height of SelectedRows
+    int64_t height = selected_rows.height();
+    os.write(reinterpret_cast<const char*>(&height), sizeof(height));
+  }
+  // the 4st field, Tensor data
+  SerializeToStream(os, selected_rows.value(), dev_ctx);
+}
+
+void DeserializeFromStream(std::istream& is, SelectedRows* selected_rows) {
+  auto tensor = *selected_rows->mutable_value();
+  {
+    // the 1st field, unit32_t version for SelectedRows
+    uint32_t version;
+    is.read(reinterpret_cast<char*>(&version), sizeof(version));
+    PADDLE_ENFORCE_EQ(version, 0U, "Only version 0 is supported");
+  }
+  {
+    // the 2st field, rows information
+    uint64_t size;
+    is.read(reinterpret_cast<char*>(&size), sizeof(size));
+    auto& rows = *selected_rows->mutable_rows();
+    rows.resize(size);
+    for (uint64_t i = 0; i < size; ++i) {
+      is.read(reinterpret_cast<char*>(&rows[i]), sizeof(int64_t));
+    }
+  }
+  {
+    // the 3st field, the height of the SelectedRows
+    int64_t height;
+    is.read(reinterpret_cast<char*>(&height), sizeof(int64_t));
+    selected_rows->set_height(height);
+  }
+  // the 4st field, tensor which contains the data
+  DeserializeFromStream(is, &tensor);
+}
+
+}  // namespace framework
 }  // namespace paddle

paddle/framework/selected_rows.h

@@ -59,5 +59,14 @@ class SelectedRows {
   int64_t height_;
 };
 
+/*
+ * Serialize/Desiralize SelectedRows to std::ostream
+ * You can pass ofstream or ostringstream to serilize to file
+ * or to a in memory string. GPU tensor will be copied to CPU.
+ */
+void SerializeToStream(std::ostream& os, const SelectedRows& selected_rows,
+                       const platform::DeviceContext& dev_ctx);
+void DeserializeFromStream(std::istream& is, SelectedRows* selected_rows);
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/selected_rows_test.cc

@@ -43,5 +43,19 @@ TEST_F(SelectedRowsTester, complete_dims) {
   ASSERT_EQ(selected_rows_->GetCompleteDims(), make_ddim({10, 100}));
 }
 
+TEST_F(SelectedRowsTester, SerializeAndDeseralize) {
+  SelectedRows dst_tensor;
+  platform::CPUDeviceContext cpu_ctx(place_);
+  std::ostringstream oss;
+
+  SerializeToStream(oss, *selected_rows_, cpu_ctx);
+
+  std::istringstream iss(oss.str());
+  DeserializeFromStream(iss, &dst_tensor);
+
+  ASSERT_EQ(selected_rows_->rows(), dst_tensor.rows());
+  ASSERT_EQ(selected_rows_->height(), dst_tensor.height());
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/shape_inference.cc

-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+/* 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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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/shape_inference.h"
 #include "grad_op_desc_maker.h"
 #include "paddle/framework/operator.h"

paddle/framework/tensor.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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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/tensor.h"
 

paddle/framework/tensor.h

@@ -20,12 +20,12 @@ limitations under the License. */
 #include <typeindex>
 #include <vector>
 
+#include "paddle/framework/data_layout.h"
 #include "paddle/framework/ddim.h"
 #include "paddle/memory/memory.h"
 #include "paddle/platform/device_context.h"
 #include "paddle/platform/enforce.h"
 #include "paddle/platform/place.h"
-#include "unsupported/Eigen/CXX11/Tensor"
 
 namespace paddle {
 
@@ -115,6 +115,10 @@ class Tensor {
 
   inline void check_memory_size() const;
 
+  inline DataLayout layout() const { return layout_; }
+
+  inline void set_layout(const DataLayout layout) { layout_ = layout; }
+
  private:
   friend class LoDTensor;
 
@@ -173,6 +177,19 @@ class Tensor {
 
   DDim dims_;
 
+  /**
+   * @brief the layout of memory block, default is NHWC.
+   *
+   * @note the memory allocation order, describe how weight/data is stored
+   *       For example, in 4-D Tensor(rank=4), there are three commonly
+   *       used layout. They are
+   *            NCHW, NHWC, CHWN.
+   *       N,C,H,W for respectively the batch size, the number of
+   *       feature maps, the height.
+   */
+
+  DataLayout layout_ = DataLayout::kNHWC;
+
   /**
    * @brief   A PlaceHolder may be shared by more than one tensor.
    *

paddle/framework/tensor.md

@@ -71,7 +71,7 @@ private:
 ```
 
 ```c++
-typedef boost::variant<GpuPlace, CpuPlace> Place;
+typedef boost::variant<CUDAPlace, CpuPlace> Place;
 typedef boost::variant<Dim<1>, Dim<2>, Dim<3>, Dim<4>, Dim<5>,
                        Dim<6>, Dim<7>, Dim<8>, Dim<9>> DDimVar;
 typedef boost::variant<

paddle/framework/tensor_impl.h

@@ -125,11 +125,11 @@ inline void* Tensor::mutable_data(platform::Place place, std::type_index type) {
           boost::get<platform::CPUPlace>(place), size, type));
     } else if (platform::is_gpu_place(place)) {
 #ifndef PADDLE_WITH_CUDA
-      PADDLE_THROW("'GPUPlace' is not supported in CPU only device.");
+      PADDLE_THROW("'CUDAPlace' is not supported in CPU only device.");
     }
 #else
-      holder_.reset(new PlaceholderImpl<platform::GPUPlace>(
-          boost::get<platform::GPUPlace>(place), size, type));
+      holder_.reset(new PlaceholderImpl<platform::CUDAPlace>(
+          boost::get<platform::CUDAPlace>(place), size, type));
     }
 #endif
     offset_ = 0;
@@ -165,6 +165,7 @@ inline Tensor Tensor::Slice(int begin_idx, int end_idx) const {
     size_t base = numel() / dims_[0];
     Tensor dst;
     dst.holder_ = holder_;
+    dst.set_layout(layout_);
     DDim dst_dims = dims_;
     dst_dims[0] = end_idx - begin_idx;
     dst.Resize(dst_dims);

paddle/framework/tensor_test.cc

@@ -15,12 +15,13 @@
 #include <gtest/gtest.h>
 #include <string>
 
+namespace framework = paddle::framework;
+namespace platform = paddle::platform;
+
 TEST(Tensor, Dims) {
-  using namespace paddle::framework;
-  using namespace paddle::platform;
-  Tensor tt;
+  framework::Tensor tt;
   tt.Resize({2, 3, 4});
-  DDim dims = tt.dims();
+  framework::DDim dims = tt.dims();
   ASSERT_EQ(arity(dims), 3);
   for (int i = 0; i < 3; ++i) {
     EXPECT_EQ(i + 2, dims[i]);
@@ -28,12 +29,12 @@ TEST(Tensor, Dims) {
 }
 
 TEST(Tensor, DataAssert) {
-  paddle::framework::Tensor src_tensor;
+  framework::Tensor src_tensor;
 
   bool caught = false;
   try {
     src_tensor.data<double>();
-  } catch (paddle::platform::EnforceNotMet err) {
+  } catch (platform::EnforceNotMet err) {
     caught = true;
     std::string msg =
         "holder_ should not be null\nTensor holds no memory. Call "
@@ -50,61 +51,65 @@ TEST(Tensor, DataAssert) {
    because Memory::Alloc() and Memory::Free() have not been ready.
 */
 TEST(Tensor, MutableData) {
-  using namespace paddle::framework;
-  using namespace paddle::platform;
   {
-    Tensor src_tensor;
+    framework::Tensor src_tensor;
     float* p1 = nullptr;
     float* p2 = nullptr;
     // initialization
-    p1 = src_tensor.mutable_data<float>(make_ddim({1, 2, 3}), CPUPlace());
+    p1 = src_tensor.mutable_data<float>(framework::make_ddim({1, 2, 3}),
+                                        platform::CPUPlace());
     EXPECT_NE(p1, nullptr);
     // set src_tensor a new dim with large size
     // momery is supposed to be re-allocated
-    p2 = src_tensor.mutable_data<float>(make_ddim({3, 4}), CPUPlace());
+    p2 = src_tensor.mutable_data<float>(framework::make_ddim({3, 4}),
+                                        platform::CPUPlace());
     EXPECT_NE(p2, nullptr);
     EXPECT_NE(p1, p2);
     // set src_tensor a new dim with same size
     // momery block is supposed to be unchanged
-    p1 = src_tensor.mutable_data<float>(make_ddim({2, 2, 3}), CPUPlace());
+    p1 = src_tensor.mutable_data<float>(framework::make_ddim({2, 2, 3}),
+                                        platform::CPUPlace());
     EXPECT_EQ(p1, p2);
     // set src_tensor a new dim with smaller size
     // momery block is supposed to be unchanged
-    p2 = src_tensor.mutable_data<float>(make_ddim({2, 2}), CPUPlace());
+    p2 = src_tensor.mutable_data<float>(framework::make_ddim({2, 2}),
+                                        platform::CPUPlace());
     EXPECT_EQ(p1, p2);
   }
 
 #ifdef PADDLE_WITH_CUDA
   {
-    Tensor src_tensor;
+    framework::Tensor src_tensor;
     float* p1 = nullptr;
     float* p2 = nullptr;
     // initialization
-    p1 = src_tensor.mutable_data<float>(make_ddim({1, 2, 3}), GPUPlace());
+    p1 = src_tensor.mutable_data<float>(framework::make_ddim({1, 2, 3}),
+                                        platform::CUDAPlace());
     EXPECT_NE(p1, nullptr);
     // set src_tensor a new dim with large size
     // momery is supposed to be re-allocated
-    p2 = src_tensor.mutable_data<float>(make_ddim({3, 4}), GPUPlace());
+    p2 = src_tensor.mutable_data<float>(framework::make_ddim({3, 4}),
+                                        platform::CUDAPlace());
     EXPECT_NE(p2, nullptr);
     EXPECT_NE(p1, p2);
     // set src_tensor a new dim with same size
     // momery block is supposed to be unchanged
-    p1 = src_tensor.mutable_data<float>(make_ddim({2, 2, 3}), GPUPlace());
+    p1 = src_tensor.mutable_data<float>(framework::make_ddim({2, 2, 3}),
+                                        platform::CUDAPlace());
     EXPECT_EQ(p1, p2);
     // set src_tensor a new dim with smaller size
     // momery block is supposed to be unchanged
-    p2 = src_tensor.mutable_data<float>(make_ddim({2, 2}), GPUPlace());
+    p2 = src_tensor.mutable_data<float>(framework::make_ddim({2, 2}),
+                                        platform::CUDAPlace());
     EXPECT_EQ(p1, p2);
   }
 #endif
 }
 
 TEST(Tensor, ShareDataWith) {
-  using namespace paddle::framework;
-  using namespace paddle::platform;
   {
-    Tensor src_tensor;
-    Tensor dst_tensor;
+    framework::Tensor src_tensor;
+    framework::Tensor dst_tensor;
     // Try to share data form uninitialized tensor
     bool caught = false;
     try {
@@ -121,16 +126,18 @@ TEST(Tensor, ShareDataWith) {
     }
     ASSERT_TRUE(caught);
 
-    src_tensor.mutable_data<int>(make_ddim({2, 3, 4}), CPUPlace());
+    src_tensor.mutable_data<int>(framework::make_ddim({2, 3, 4}),
+                                 platform::CPUPlace());
     dst_tensor.ShareDataWith(src_tensor);
     ASSERT_EQ(src_tensor.data<int>(), dst_tensor.data<int>());
   }
 
 #ifdef PADDLE_WITH_CUDA
   {
-    Tensor src_tensor;
-    Tensor dst_tensor;
-    src_tensor.mutable_data<int>(make_ddim({2, 3, 4}), GPUPlace());
+    framework::Tensor src_tensor;
+    framework::Tensor dst_tensor;
+    src_tensor.mutable_data<int>(framework::make_ddim({2, 3, 4}),
+                                 platform::CUDAPlace());
     dst_tensor.ShareDataWith(src_tensor);
     ASSERT_EQ(src_tensor.data<int>(), dst_tensor.data<int>());
   }
@@ -138,13 +145,12 @@ TEST(Tensor, ShareDataWith) {
 }
 
 TEST(Tensor, Slice) {
-  using namespace paddle::framework;
-  using namespace paddle::platform;
   {
-    Tensor src_tensor;
-    src_tensor.mutable_data<int>(make_ddim({5, 3, 4}), CPUPlace());
-    Tensor slice_tensor = src_tensor.Slice(1, 3);
-    DDim slice_dims = slice_tensor.dims();
+    framework::Tensor src_tensor;
+    src_tensor.mutable_data<int>(framework::make_ddim({5, 3, 4}),
+                                 platform::CPUPlace());
+    framework::Tensor slice_tensor = src_tensor.Slice(1, 3);
+    framework::DDim slice_dims = slice_tensor.dims();
     ASSERT_EQ(arity(slice_dims), 3);
     EXPECT_EQ(slice_dims[0], 2);
     EXPECT_EQ(slice_dims[1], 3);
@@ -153,11 +159,12 @@ TEST(Tensor, Slice) {
     uintptr_t src_data_address =
         reinterpret_cast<uintptr_t>(src_tensor.data<int>());
     uintptr_t src_mutable_data_address = reinterpret_cast<uintptr_t>(
-        src_tensor.mutable_data<int>(src_tensor.dims(), CPUPlace()));
+        src_tensor.mutable_data<int>(src_tensor.dims(), platform::CPUPlace()));
     uintptr_t slice_data_address =
         reinterpret_cast<uintptr_t>(slice_tensor.data<int>());
-    uintptr_t slice_mutable_data_address = reinterpret_cast<uintptr_t>(
-        slice_tensor.mutable_data<int>(slice_tensor.dims(), CPUPlace()));
+    uintptr_t slice_mutable_data_address =
+        reinterpret_cast<uintptr_t>(slice_tensor.mutable_data<int>(
+            slice_tensor.dims(), platform::CPUPlace()));
     EXPECT_EQ(src_data_address, src_mutable_data_address);
     EXPECT_EQ(slice_data_address, slice_mutable_data_address);
     EXPECT_EQ(src_data_address + 3 * 4 * 1 * sizeof(int), slice_data_address);
@@ -165,22 +172,25 @@ TEST(Tensor, Slice) {
 
 #ifdef PADDLE_WITH_CUDA
   {
-    Tensor src_tensor;
-    src_tensor.mutable_data<double>(make_ddim({6, 9}), GPUPlace());
-    Tensor slice_tensor = src_tensor.Slice(2, 6);
-    DDim slice_dims = slice_tensor.dims();
+    framework::Tensor src_tensor;
+    src_tensor.mutable_data<double>(framework::make_ddim({6, 9}),
+                                    platform::CUDAPlace());
+    framework::Tensor slice_tensor = src_tensor.Slice(2, 6);
+    framework::DDim slice_dims = slice_tensor.dims();
     ASSERT_EQ(arity(slice_dims), 2);
     EXPECT_EQ(slice_dims[0], 4);
     EXPECT_EQ(slice_dims[1], 9);
 
     uintptr_t src_data_address =
         reinterpret_cast<uintptr_t>(src_tensor.data<double>());
-    uintptr_t src_mutable_data_address = reinterpret_cast<uintptr_t>(
-        src_tensor.mutable_data<double>(src_tensor.dims(), GPUPlace()));
+    uintptr_t src_mutable_data_address =
+        reinterpret_cast<uintptr_t>(src_tensor.mutable_data<double>(
+            src_tensor.dims(), platform::CUDAPlace()));
     uintptr_t slice_data_address =
         reinterpret_cast<uintptr_t>(slice_tensor.data<double>());
-    uintptr_t slice_mutable_data_address = reinterpret_cast<uintptr_t>(
-        slice_tensor.mutable_data<double>(slice_tensor.dims(), GPUPlace()));
+    uintptr_t slice_mutable_data_address =
+        reinterpret_cast<uintptr_t>(slice_tensor.mutable_data<double>(
+            slice_tensor.dims(), platform::CUDAPlace()));
     EXPECT_EQ(src_data_address, src_mutable_data_address);
     EXPECT_EQ(slice_data_address, slice_mutable_data_address);
     EXPECT_EQ(src_data_address + 9 * 2 * sizeof(double), slice_data_address);
@@ -189,14 +199,19 @@ TEST(Tensor, Slice) {
 }
 
 TEST(Tensor, ReshapeToMatrix) {
-  using namespace paddle::framework;
-  using namespace paddle::platform;
-  Tensor src;
-  int* src_ptr = src.mutable_data<int>({2, 3, 4, 9}, CPUPlace());
+  framework::Tensor src;
+  int* src_ptr = src.mutable_data<int>({2, 3, 4, 9}, platform::CPUPlace());
   for (int i = 0; i < 2 * 3 * 4 * 9; ++i) {
     src_ptr[i] = i;
   }
-  Tensor res = ReshapeToMatrix(src, 2);
+  framework::Tensor res = framework::ReshapeToMatrix(src, 2);
   ASSERT_EQ(res.dims()[0], 2 * 3);
   ASSERT_EQ(res.dims()[1], 4 * 9);
 }
+
+TEST(Tensor, Layout) {
+  framework::Tensor src;
+  ASSERT_EQ(src.layout(), framework::DataLayout::kNHWC);
+  src.set_layout(framework::DataLayout::kAnyLayout);
+  ASSERT_EQ(src.layout(), framework::DataLayout::kAnyLayout);
+}

paddle/framework/tensor_util.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+   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/tensor_util.h"
+
+namespace paddle {
+namespace framework {
+template <typename Predicate, typename DevCtx>
+struct AnyDTypeVisitor {
+  Predicate predicate_;
+  const Tensor& tensor_;
+  const DevCtx& ctx_;
+  Tensor* out_;
+
+  AnyDTypeVisitor(Predicate predicate, const Tensor& tensor, const DevCtx& ctx,
+                  Tensor* out)
+      : predicate_(predicate), tensor_(tensor), ctx_(ctx), out_(out) {}
+
+  template <typename T>
+  void operator()() const {
+    auto t = EigenVector<T>::Flatten(tensor_);
+    auto o = EigenScalar<bool>::From(*out_);
+    // return any of predicate_(t) is true.
+    o.device(*ctx_.eigen_device()) = predicate_(t).any();
+  }
+};
+
+template <typename Predicate, typename DevCtx>
+inline void AnyImpl(Predicate predicate, const framework::Tensor& tensor,
+                    const DevCtx& ctx, framework::Tensor* out) {
+  VisitDataType(ToDataType(tensor.type()), AnyDTypeVisitor<Predicate, DevCtx>(
+                                               predicate, tensor, ctx, out));
+}
+
+template <typename Predicate>
+struct AnyVisitor : public boost::static_visitor<bool> {
+  const framework::Tensor& tensor_;
+  Predicate predicate_;
+
+  AnyVisitor(const framework::Tensor& tensor, Predicate predicate)
+      : tensor_(tensor), predicate_(std::move(predicate)) {}
+
+  template <typename Place>
+  bool operator()(const Place& place) const {
+    framework::Tensor out;
+    out.Resize({1});
+    out.mutable_data<bool>(place);
+    auto* ctx = platform::DeviceContextPool::Instance().GetByPlace(place);
+    AnyImpl(predicate_, tensor_, *ctx, &out);
+    return this->GetResult(out, place);
+  }
+
+  bool GetResult(const framework::Tensor& out,
+                 const platform::CUDAPlace& gpu) const {
+    platform::CPUPlace cpu;
+    framework::Tensor tmp;
+    tmp.Resize({1});
+    tmp.mutable_data<bool>(cpu);
+    auto gpuctx = platform::DeviceContextPool::Instance().Get(gpu);
+    gpuctx->Wait();
+    CopyFrom(out, cpu, *gpuctx, &tmp);
+    gpuctx->Wait();
+    return GetResult(tmp, cpu);
+  }
+
+  bool GetResult(const framework::Tensor& out,
+                 const platform::CPUPlace& cpu) const {
+    return *out.data<bool>();
+  }
+};
+
+template <typename Predicate>
+inline bool Any(const framework::Tensor& tensor, Predicate predicate) {
+  AnyVisitor<Predicate> visitor(tensor, predicate);
+  auto place = tensor.place();
+  return platform::VisitPlace(place, visitor);
+}
+
+struct HasNANPredicate {
+  template <typename T>
+  auto operator()(const T& eigen_vec) const
+      -> decltype(std::declval<T>().isnan()) {
+    // Cast eigen_vector to vector of bool. true if is inf.
+    return eigen_vec.isnan();
+  }
+};
+
+bool HasNAN(const framework::Tensor& tensor) {
+  HasNANPredicate predicate;
+  return Any(tensor, predicate);
+}
+
+struct HasInfPredicate {
+  template <typename T>
+  auto operator()(const T& eigen_vec) const
+      -> decltype(std::declval<T>().isinf()) {
+    // Cast eigen_vector to vector of bool. true if is inf.
+    return eigen_vec.isinf();
+  }
+};
+
+bool HasInf(const framework::Tensor& tensor) {
+  HasInfPredicate predicate;
+  return Any(tensor, predicate);
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/tensor_util.cu

+./tensor_util.cc
\ No newline at end of file

paddle/framework/tensor_util.h

 /* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
 
-   Licensed under the Apache License, Version 2.0 (the "License");
-   you may not use this file except in compliance with the License.
-   You may obtain a copy of the License at
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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/data_type.h"
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/framework.pb.h"
 #include "paddle/framework/tensor.h"
+#include "paddle/platform/device_context.h"
 
 namespace paddle {
 namespace framework {
@@ -33,6 +37,7 @@ inline void CopyFrom(const Tensor& src, const platform::Place& dst_place,
   src.check_memory_size();
 
   dst->Resize(src.dims());
+  dst->set_layout(src.layout());
   auto src_place = src.place();
   auto src_ptr = src.data<void>();
 
@@ -47,11 +52,11 @@ inline void CopyFrom(const Tensor& src, const platform::Place& dst_place,
 #ifdef PADDLE_WITH_CUDA
   else if (platform::is_gpu_place(src_place) &&  // NOLINT
            platform::is_cpu_place(dst_place)) {
-    auto src_gpu_place = boost::get<platform::GPUPlace>(src_place);
+    auto src_gpu_place = boost::get<platform::CUDAPlace>(src_place);
     auto dst_cpu_place = boost::get<platform::CPUPlace>(dst_place);
     auto ctx_place = ctx.GetPlace();
     PADDLE_ENFORCE(platform::is_gpu_place(ctx_place));
-    auto ctx_gpu_place = boost::get<platform::GPUPlace>(ctx_place);
+    auto ctx_gpu_place = boost::get<platform::CUDAPlace>(ctx_place);
     PADDLE_ENFORCE_EQ(src_gpu_place, ctx_gpu_place);
     memory::Copy(
         dst_cpu_place, dst_ptr, src_gpu_place, src_ptr, size,
@@ -59,21 +64,21 @@ inline void CopyFrom(const Tensor& src, const platform::Place& dst_place,
   } else if (platform::is_cpu_place(src_place) &&
              platform::is_gpu_place(dst_place)) {
     auto src_cpu_place = boost::get<platform::CPUPlace>(src_place);
-    auto dst_gpu_place = boost::get<platform::GPUPlace>(dst_place);
+    auto dst_gpu_place = boost::get<platform::CUDAPlace>(dst_place);
     auto ctx_place = ctx.GetPlace();
     PADDLE_ENFORCE(platform::is_gpu_place(ctx_place));
-    auto ctx_gpu_place = boost::get<platform::GPUPlace>(ctx_place);
+    auto ctx_gpu_place = boost::get<platform::CUDAPlace>(ctx_place);
     PADDLE_ENFORCE_EQ(dst_gpu_place, ctx_gpu_place);
     memory::Copy(
         dst_gpu_place, dst_ptr, src_cpu_place, src_ptr, size,
         reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream());
   } else if (platform::is_gpu_place(src_place) &&
              platform::is_gpu_place(dst_place)) {
-    auto src_gpu_place = boost::get<platform::GPUPlace>(src_place);
-    auto dst_gpu_place = boost::get<platform::GPUPlace>(dst_place);
+    auto src_gpu_place = boost::get<platform::CUDAPlace>(src_place);
+    auto dst_gpu_place = boost::get<platform::CUDAPlace>(dst_place);
     auto ctx_place = ctx.GetPlace();
     PADDLE_ENFORCE(platform::is_gpu_place(ctx_place));
-    auto ctx_gpu_place = boost::get<platform::GPUPlace>(ctx_place);
+    auto ctx_gpu_place = boost::get<platform::CUDAPlace>(ctx_place);
     PADDLE_ENFORCE_EQ(src_gpu_place, ctx_gpu_place);
     memory::Copy(
         dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size,
@@ -82,6 +87,29 @@ inline void CopyFrom(const Tensor& src, const platform::Place& dst_place,
 #endif
 }
 
+/**
+ * @brief CopyFrom support CPU <-> CPU
+ */
+inline void CopyFrom(const Tensor& src, const platform::Place& dst_place,
+                     Tensor* dst) {
+  src.check_memory_size();
+  dst->Resize(src.dims());
+  dst->set_layout(src.layout());
+
+  auto src_place = src.place();
+  auto src_ptr = src.data<void>();
+
+  auto dst_ptr = dst->mutable_data(dst_place, src.type());
+
+  auto size = src.numel() * SizeOfType(src.type());
+
+  PADDLE_ENFORCE(platform::is_cpu_place(src_place) &&
+                 platform::is_cpu_place(dst_place));
+
+  memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
+               boost::get<platform::CPUPlace>(src_place), src_ptr, size);
+}
+
 /**
  * @brief   Copy the content of an external vector to a tensor.
  *
@@ -108,13 +136,28 @@ inline void CopyFromVector(const std::vector<T>& src,
 #ifdef PADDLE_WITH_CUDA
   else if (platform::is_gpu_place(dst_place)) {  // NOLINT
     memory::Copy(
-        boost::get<platform::GPUPlace>(dst_place), dst_ptr, src_place, src_ptr,
+        boost::get<platform::CUDAPlace>(dst_place), dst_ptr, src_place, src_ptr,
         size,
         reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream());
   }
 #endif
 }
 
+/**
+ * @brief CopyFromVector CPU vector -> CPU Tensor
+ */
+template <typename T>
+inline void CopyFromVector(const std::vector<T>& src, Tensor* dst) {
+  platform::CPUPlace dst_place = platform::CPUPlace();
+  auto src_ptr = static_cast<const void*>(src.data());
+  platform::CPUPlace src_place;
+  dst->Resize({static_cast<int64_t>(src.size())});
+  auto dst_ptr = static_cast<void*>(dst->mutable_data<T>(dst_place));
+  auto size = src.size() * sizeof(T);
+
+  memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
+}
+
 /**
  * @brief   Copy the content of a tensor to a vector
  *
@@ -141,12 +184,134 @@ inline void CopyToVector(const Tensor& src, const platform::DeviceContext& ctx,
 #ifdef PADDLE_WITH_CUDA
   else if (platform::is_gpu_place(src.place())) {  // NOLINT
     memory::Copy(
-        dst_place, dst_ptr, boost::get<platform::GPUPlace>(src.place()),
+        dst_place, dst_ptr, boost::get<platform::CUDAPlace>(src.place()),
         src_ptr, size,
         reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream());
   }
 #endif
 }
 
+/**
+ * @brief CopyToVector CPUTensor <-> CPU Vector
+ */
+template <typename T>
+inline void CopyToVector(const Tensor& src, std::vector<T>* dst) {
+  auto src_ptr = static_cast<const void*>(src.data<T>());
+  auto size = src.numel() * sizeof(T);
+
+  platform::CPUPlace dst_place;
+  dst->resize(src.numel());
+  auto dst_ptr = static_cast<void*>(dst->data());
+
+  PADDLE_ENFORCE(platform::is_cpu_place(src.place()));
+
+  memory::Copy(dst_place, dst_ptr, boost::get<platform::CPUPlace>(src.place()),
+               src_ptr, size);
+}
+
+// Returns true if a tensor contains NAN, i.e., Not A Number.
+bool HasNAN(const framework::Tensor& tensor);
+
+// Returns true if a tensor contains Inf, i.e., Infinity.
+bool HasInf(const framework::Tensor& tensor);
+
+inline void SerializeToStream(std::ostream& os, const Tensor& tensor,
+                              const platform::DeviceContext& dev_ctx) {
+  // TODO(typhoonzero): serialize to ostream
+  {  // the 1st field, uint32_t version
+    constexpr uint32_t version = 0;
+    os.write(reinterpret_cast<const char*>(&version), sizeof(version));
+  }
+  {  // the 2nd field, tensor description
+     // int32_t  size
+     // void*    protobuf message
+    proto::TensorDesc desc;
+    desc.set_data_type(framework::ToDataType(tensor.type()));
+    auto dims = framework::vectorize(tensor.dims());
+    auto* pb_dims = desc.mutable_dims();
+    pb_dims->Resize(static_cast<int>(dims.size()), 0);
+    std::copy(dims.begin(), dims.end(), pb_dims->begin());
+    int32_t size = desc.ByteSize();
+    os.write(reinterpret_cast<const char*>(&size), sizeof(size));
+    auto out = desc.SerializeAsString();
+    os.write(out.data(), size);
+  }
+  {  // the 3rd field, tensor data
+    uint64_t size = tensor.memory_size();
+    auto* data_ptr = tensor.data<void>();
+    PADDLE_ENFORCE(size < std::numeric_limits<std::streamsize>::max(),
+                   "Index overflow when writing tensor");
+    if (platform::is_gpu_place(tensor.place())) {
+#ifdef PADDLE_WITH_CUDA
+      constexpr size_t kBufSize = 1024 * 1024 * 64;  // 64MB
+      std::unique_ptr<char[]> buf(new char[kBufSize]);
+      auto& gpu_dev_ctx =
+          static_cast<const platform::CUDADeviceContext&>(dev_ctx);
+      platform::CPUPlace cpu;
+      uintptr_t data = reinterpret_cast<uintptr_t>(data_ptr);
+      while (size != 0) {
+        size_t size_to_write = std::min(kBufSize, static_cast<size_t>(size));
+        memory::Copy(cpu, buf.get(),
+                     boost::get<platform::CUDAPlace>(tensor.place()),
+                     reinterpret_cast<const void*>(data), size_to_write,
+                     gpu_dev_ctx.stream());
+        gpu_dev_ctx.Wait();
+        os.write(buf.get(), size_to_write);
+        data += size_to_write;
+        size -= size_to_write;
+      }
+#else
+      PADDLE_THROW("Unexpected branch");
+#endif
+    } else {
+      os.write(static_cast<const char*>(data_ptr),
+               static_cast<std::streamsize>(size));
+    }
+  }
+}
+
+inline void DeserializeFromStream(std::istream& is, Tensor* tensor) {
+  uint32_t version;
+  is.read(reinterpret_cast<char*>(&version), sizeof(version));
+  PADDLE_ENFORCE_EQ(version, 0U, "Only version 0 is supported");
+  proto::TensorDesc desc;
+  {  // int32_t size
+     // proto buffer
+    int32_t size;
+    is.read(reinterpret_cast<char*>(&size), sizeof(size));
+    std::unique_ptr<char[]> buf(new char[size]);
+    is.read(reinterpret_cast<char*>(buf.get()), size);
+    PADDLE_ENFORCE(desc.ParseFromArray(buf.get(), size),
+                   "Cannot parse tensor desc");
+  }
+  {  // read tensor
+    std::vector<int64_t> dims;
+    dims.reserve(static_cast<size_t>(desc.dims().size()));
+    std::copy(desc.dims().begin(), desc.dims().end(), std::back_inserter(dims));
+    tensor->Resize(framework::make_ddim(dims));
+
+    void* buf;
+    platform::Place cpu = platform::CPUPlace();
+    // TODO(Yancey1989): use VisiterDataType instead of DataType switch
+    switch (desc.data_type()) {
+      case proto::FP32:
+        buf = tensor->mutable_data<float>(cpu);
+        break;
+      case proto::FP64:
+        buf = tensor->mutable_data<double>(cpu);
+        break;
+      case proto::INT32:
+        buf = tensor->mutable_data<int>(cpu);
+        break;
+      case proto::INT64:
+        buf = tensor->mutable_data<int64_t>(cpu);
+        break;
+      default:
+        PADDLE_THROW("DataType %d not supported", desc.data_type());
+    }
+    is.read(static_cast<char*>(buf), tensor->memory_size());
+  }
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/tensor_util_test.cc

@@ -13,10 +13,12 @@
 
 #include "paddle/framework/tensor_util.h"
 #include <gtest/gtest.h>
+#include <cmath>
 #include <string>
 
 namespace paddle {
 namespace framework {
+
 TEST(CopyFrom, Tensor) {
   Tensor src_tensor;
   Tensor dst_tensor;
@@ -27,9 +29,10 @@ TEST(CopyFrom, Tensor) {
 
   int arr[9] = {1, 2, 3, 4, 5, 6, 7, 8, 9};
   memcpy(src_ptr, arr, 9 * sizeof(int));
+  src_tensor.set_layout(DataLayout::kAnyLayout);
 
   auto cpu_place = new platform::CPUPlace();
-  CopyFrom(src_tensor, *cpu_place, cpu_ctx, &dst_tensor);
+  CopyFrom(src_tensor, *cpu_place, &dst_tensor);
 
   const int* dst_ptr = dst_tensor.data<int>();
   ASSERT_NE(src_ptr, dst_ptr);
@@ -37,14 +40,18 @@ TEST(CopyFrom, Tensor) {
     EXPECT_EQ(src_ptr[i], dst_ptr[i]);
   }
 
+  EXPECT_TRUE(dst_tensor.layout() == src_tensor.layout());
+
   Tensor slice_tensor = src_tensor.Slice(1, 2);
-  CopyFrom(slice_tensor, *cpu_place, cpu_ctx, &dst_tensor);
+  CopyFrom(slice_tensor, *cpu_place, &dst_tensor);
   const int* slice_ptr = slice_tensor.data<int>();
   dst_ptr = dst_tensor.data<int>();
   ASSERT_NE(dst_ptr, slice_ptr);
   for (size_t i = 0; i < 3; ++i) {
     EXPECT_EQ(dst_ptr[i], slice_ptr[i]);
   }
+  EXPECT_TRUE(dst_tensor.layout() == src_tensor.layout());
+
 #ifdef PADDLE_WITH_CUDA
   {
     Tensor src_tensor;
@@ -58,7 +65,7 @@ TEST(CopyFrom, Tensor) {
     memcpy(src_ptr, arr, 9 * sizeof(int));
 
     // CPU Tensor to GPU Tensor
-    auto gpu_place = new platform::GPUPlace(0);
+    auto gpu_place = new platform::CUDAPlace(0);
     platform::CUDADeviceContext gpu_ctx(*gpu_place);
     CopyFrom(src_tensor, *gpu_place, gpu_ctx, &gpu_tensor);
 
@@ -90,6 +97,8 @@ TEST(CopyFrom, Tensor) {
     for (size_t i = 0; i < 3; ++i) {
       EXPECT_EQ(dst_ptr[i], slice_ptr[i]);
     }
+
+    EXPECT_TRUE(dst_tensor.layout() == src_tensor.layout());
   }
 #endif
 }
@@ -104,8 +113,7 @@ TEST(CopyFromVector, Tensor) {
     // Copy to CPU Tensor
     cpu_tensor.Resize(make_ddim({3, 3}));
     auto cpu_place = new paddle::platform::CPUPlace();
-    CPUDeviceContext cpu_ctx(*cpu_place);
-    CopyFromVector<int>(src_vec, cpu_ctx, &cpu_tensor);
+    CopyFromVector<int>(src_vec, &cpu_tensor);
 
     // Compare Tensors
     const int* cpu_ptr = cpu_tensor.data<int>();
@@ -117,7 +125,7 @@ TEST(CopyFromVector, Tensor) {
 
     src_vec.erase(src_vec.begin(), src_vec.begin() + 5);
     cpu_tensor.Resize(make_ddim({2, 2}));
-    CopyFromVector<int>(src_vec, cpu_ctx, &cpu_tensor);
+    CopyFromVector<int>(src_vec, &cpu_tensor);
     cpu_ptr = cpu_tensor.data<int>();
     src_ptr = src_vec.data();
     ASSERT_NE(src_ptr, cpu_ptr);
@@ -143,7 +151,7 @@ TEST(CopyFromVector, Tensor) {
 
     // Copy to GPUTensor
     gpu_tensor.Resize(make_ddim({3, 3}));
-    auto gpu_place = new paddle::platform::GPUPlace();
+    auto gpu_place = new paddle::platform::CUDAPlace();
     CUDADeviceContext gpu_ctx(*gpu_place);
     CopyFromVector<int>(src_vec, gpu_ctx, &gpu_tensor);
     // Copy from GPU to CPU tensor for comparison
@@ -198,9 +206,8 @@ TEST(CopyToVector, Tensor) {
     }
 
     CPUPlace place;
-    CPUDeviceContext cpu_ctx(place);
     std::vector<int> dst;
-    CopyToVector<int>(src, cpu_ctx, &dst);
+    CopyToVector<int>(src, &dst);
 
     for (int i = 0; i < 3 * 3; ++i) {
       EXPECT_EQ(src_ptr[i], dst[i]);
@@ -210,7 +217,7 @@ TEST(CopyToVector, Tensor) {
   {
     std::vector<int> src_vec = {1, 2, 3, 4, 5, 6, 7, 8, 9};
     Tensor gpu_tensor;
-    GPUPlace place;
+    CUDAPlace place;
     CUDADeviceContext gpu_ctx(place);
     CopyFromVector<int>(src_vec, gpu_ctx, &gpu_tensor);
 
@@ -224,5 +231,78 @@ TEST(CopyToVector, Tensor) {
 #endif
 }
 
+TEST(HasNAN, CPU) {
+  using namespace paddle::framework;
+  using namespace paddle::platform;
+  Tensor src;
+  float* buf = src.mutable_data<float>({3}, CPUPlace());
+  buf[0] = 0.0;
+  buf[1] = NAN;
+  buf[2] = 0.0;
+
+  ASSERT_TRUE(HasNAN(src));
+}
+
+TEST(HasInf, CPU) {
+  using namespace paddle::framework;
+  using namespace paddle::platform;
+  Tensor src;
+  double* buf = src.mutable_data<double>({3}, CPUPlace());
+  buf[0] = 1.0;
+  buf[1] = INFINITY;
+  buf[2] = 0.0;
+  ASSERT_TRUE(HasInf(src));
+}
+
+TEST(Tensor, SerializeAndDeserialize) {
+  framework::Tensor src_tensor;
+  int array[6] = {1, 2, 3, 4, 5, 6};
+  src_tensor.Resize({2, 3});
+  int* src_ptr = src_tensor.mutable_data<int>(platform::CPUPlace());
+  for (int i = 0; i < 6; ++i) {
+    src_ptr[i] = array[i];
+  }
+  {
+    framework::Tensor dst_tensor;
+    auto place = new platform::CPUPlace();
+    platform::CPUDeviceContext cpu_ctx(*place);
+    std::ostringstream oss;
+    SerializeToStream(oss, src_tensor, cpu_ctx);
+
+    std::istringstream iss(oss.str());
+    DeserializeFromStream(iss, &dst_tensor);
+    int* dst_ptr = dst_tensor.mutable_data<int>(platform::CPUPlace());
+    for (int i = 0; i < 5; ++i) {
+      ASSERT_EQ(dst_ptr[i], array[i]);
+    }
+    delete place;
+  }
+#ifdef PADDLE_WITH_CUDA
+  {
+    Tensor gpu_tensor;
+    gpu_tensor.Resize({2, 3});
+    Tensor dst_tensor;
+
+    auto gpu_place = new platform::CUDAPlace();
+    platform::CUDADeviceContext gpu_ctx(*gpu_place);
+
+    CopyFrom(src_tensor, *gpu_place, gpu_ctx, &gpu_tensor);
+
+    std::ostringstream oss;
+    SerializeToStream(oss, gpu_tensor, gpu_ctx);
+
+    std::istringstream iss(oss.str());
+    DeserializeFromStream(iss, &dst_tensor);
+
+    int* dst_ptr = dst_tensor.mutable_data<int>(platform::CPUPlace());
+    for (int i = 0; i < 6; ++i) {
+      ASSERT_EQ(dst_ptr[i], array[i]);
+    }
+
+    delete gpu_place;
+  }
+#endif
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/tensor_util_test.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+   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 "paddle/framework/tensor_util.h"
+#include "paddle/platform/device_context.h"
+#include "paddle/platform/place.h"
+
+namespace paddle {
+namespace framework {
+
+static __global__ void FillNAN(float* buf) {
+  buf[0] = 0.0;
+  buf[1] = 0.1;
+  buf[2] = NAN;
+}
+static __global__ void FillInf(float* buf) {
+  buf[0] = 0.0;
+  buf[1] = INFINITY;
+  buf[2] = 0.5;
+}
+
+TEST(HasNAN, GPU) {
+  Tensor tensor;
+  platform::CUDAPlace gpu(0);
+  auto& pool = platform::DeviceContextPool::Instance();
+  auto* cuda_ctx = pool.GetByPlace(gpu);
+  float* buf = tensor.mutable_data<float>({3}, gpu);
+  FillNAN<<<1, 1, 0, cuda_ctx->stream()>>>(buf);
+  cuda_ctx->Wait();
+  ASSERT_TRUE(HasNAN(tensor));
+}
+
+TEST(HasInf, GPU) {
+  Tensor tensor;
+  platform::CUDAPlace gpu(0);
+  auto& pool = platform::DeviceContextPool::Instance();
+  auto* cuda_ctx = pool.GetByPlace(gpu);
+  float* buf = tensor.mutable_data<float>({3}, gpu);
+  FillInf<<<1, 1, 0, cuda_ctx->stream()>>>(buf);
+  cuda_ctx->Wait();
+  ASSERT_TRUE(HasInf(tensor));
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/threadpool.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/framework/threadpool.h"
+
+namespace paddle {
+namespace framework {
+
+std::unique_ptr<ThreadPool> ThreadPool::threadpool(nullptr);
+std::once_flag ThreadPool::init_flag;
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/threadpool.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 <condition_variable>
+#include <functional>
+#include <future>
+#include <mutex>
+#include <queue>
+#include <thread>
+
+#include "paddle/platform/enforce.h"
+
+namespace paddle {
+namespace framework {
+
+class ThreadPool {
+ public:
+  typedef std::packaged_task<void()> Task;
+  typedef std::function<void()> Fun;
+
+  /**
+   * @brief   Get a instance of threadpool, the thread number will
+   *          be specified as the number of hardware thread contexts
+   */
+  static ThreadPool* GetInstance() {
+    std::call_once(init_flag, &ThreadPool::Init);
+    return threadpool.get();
+  }
+
+  ~ThreadPool() {
+    {
+      // notify all threads to stop running
+      running_ = false;
+      scheduled_.notify_all();
+    }
+
+    for (auto& t : threads_) {
+      t->join();
+      t.reset(nullptr);
+    }
+  }
+
+  int GetNumThreads() const { return num_threads_; }
+
+  int GetAvailable() {
+    std::unique_lock<std::mutex> lock(mutex_);
+    return available_;
+  }
+
+  /**
+   * @brief   Push a function to the queue, and will be scheduled and
+   *          executed if a thread is available.
+   * @param[in] Task, will be pushed to the task queue.
+   * @return    std::future<void>, we could wait for the task finished by
+   *            f.wait().
+   */
+  std::future<void> Run(const Fun& fn) {
+    std::unique_lock<std::mutex> lock(mutex_);
+    Task task(std::bind(fn));
+    std::future<void> f = task.get_future();
+    tasks_.push(std::move(task));
+    lock.unlock();
+    scheduled_.notify_one();
+    return f;
+  }
+
+  /**
+   * @brief   Wait until all the tasks are completed.
+   */
+  void Wait() {
+    std::unique_lock<std::mutex> lock(mutex_);
+    completed_.wait(lock, [=] { return Done() == true; });
+  }
+
+ private:
+  DISABLE_COPY_AND_ASSIGN(ThreadPool);
+
+  explicit ThreadPool(int num_threads)
+      : num_threads_(num_threads), available_(num_threads), running_(true) {
+    threads_.resize(num_threads);
+    for (auto& thread : threads_) {
+      // TODO(Yancey1989): binding the thread on the specify CPU number
+      thread.reset(new std::thread(std::bind(&ThreadPool::TaskLoop, this)));
+    }
+  }
+
+  /**
+   * @brief   If the task queue is empty and avaialbe
+   *          is equal to the number of threads, means that
+   *          all tasks are completed.
+   *
+   *          Note: this function is not thread-safe.
+   *
+   * @return true if all tasks are completed.
+   */
+  bool Done() { return tasks_.empty() && available_ == num_threads_; }
+
+  void TaskLoop() {
+    while (running_) {
+      std::unique_lock<std::mutex> lock(mutex_);
+      scheduled_.wait(lock, [=] { return !tasks_.empty() || !running_; });
+
+      if (!running_) {
+        break;
+      }
+      // pop a task from the task queue
+      auto task = std::move(tasks_.front());
+      tasks_.pop();
+
+      --available_;
+      lock.unlock();
+
+      // run the task
+      task();
+
+      {
+        std::unique_lock<std::mutex> lock(mutex_);
+        ++available_;
+        if (Done()) {
+          completed_.notify_all();
+        }
+      }
+    }
+  }
+
+  static void Init() {
+    if (threadpool.get() == nullptr) {
+      // TODO(Yancey1989): specify the max threads number
+      int num_threads = std::thread::hardware_concurrency();
+      PADDLE_ENFORCE_GT(num_threads, 0);
+      threadpool.reset(new ThreadPool(num_threads));
+    }
+  }
+
+ private:
+  static std::unique_ptr<ThreadPool> threadpool;
+  static std::once_flag init_flag;
+
+  int num_threads_;
+  int available_;
+  bool running_;
+  std::queue<Task> tasks_;
+  std::vector<std::unique_ptr<std::thread>> threads_;
+  std::mutex mutex_;
+  std::condition_variable scheduled_;
+  std::condition_variable completed_;
+};
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/threadpool_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 <gtest/gtest.h>
+#include <atomic>
+
+#include "threadpool.h"
+
+namespace framework = paddle::framework;
+
+void do_sum(framework::ThreadPool* pool, std::atomic<int>& sum, int cnt) {
+  std::vector<std::future<void>> fs;
+  for (int i = 0; i < cnt; ++i) {
+    auto f = pool->Run([&sum]() { sum.fetch_add(1); });
+    fs.push_back(std::move(f));
+  }
+  for (auto& f : fs) {
+    f.wait();
+  }
+}
+
+TEST(ThreadPool, ConcurrentInit) {
+  framework::ThreadPool* pool;
+  int n = 50;
+  std::vector<std::thread> threads;
+  for (int i = 0; i < n; ++i) {
+    std::thread t([&pool]() { pool = framework::ThreadPool::GetInstance(); });
+    threads.push_back(std::move(t));
+  }
+  for (auto& t : threads) {
+    t.join();
+  }
+}
+
+TEST(ThreadPool, ConcurrentRun) {
+  framework::ThreadPool* pool = framework::ThreadPool::GetInstance();
+  std::atomic<int> sum(0);
+  std::vector<std::thread> threads;
+  int n = 50;
+  // sum = (n * (n + 1)) / 2
+  for (int i = 1; i <= n; ++i) {
+    std::thread t(do_sum, pool, std::ref(sum), i);
+    threads.push_back(std::move(t));
+  }
+  for (auto& t : threads) {
+    t.join();
+  }
+  pool->Wait();
+  EXPECT_EQ(sum, ((n + 1) * n) / 2);
+}

paddle/framework/type_defs.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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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 <functional>

paddle/framework/var_desc.cc

@@ -74,7 +74,7 @@ const proto::TensorDesc &VarDesc::tensor_desc() const {
     case proto::VarDesc::LOD_TENSOR_ARRAY:
       return desc_.tensor_array().tensor();
     default:
-      PADDLE_THROW("Unexpected branch.");
+      PADDLE_THROW("The type of var '", this->Name(), "' is unsupported.");
   }
 }
 

paddle/framework/var_type.h

-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+/* 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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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/framework.pb.h"

paddle/framework/var_type_inference.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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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/type_defs.h"

paddle/framework/var_type_inference_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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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/var_type_inference.h"
 #include "gtest/gtest.h"

paddle/function/GemmConvOp.cpp

@@ -126,14 +126,165 @@ public:
       inputData += inputChannels * inputHeight * inputWidth;
       outputData += outputChannels * outputHeight * outputWidth;
     }
+  }
+};
+
 #ifdef PADDLE_MOBILE_INFERENCE
-    if (Device == DEVICE_TYPE_CPU) {
-      memory_.reset();
+
+/*
+ * \brief Forward calculation of convolution, optimized for mobile.
+ */
+template <DeviceType Device>
+class GemmConvMobileFunction : 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);
+    // TODO(hedaoyuan): Need to define some index macros,
+    // to avoid useing 0 and 1.
+    const TensorShape& input = inputs[0].shape();
+    const TensorShape& filter = inputs[1].shape();
+    const TensorShape& output = outputs[0].shape();
+
+    real beta;
+    if (outputs[0].getArgType() == ADD_TO) {
+      beta = 1.0;
+    } else {
+      beta = 0.0;
     }
-#endif
+
+    size_t batchSize = input[0];
+    size_t inputChannels = input[1];
+    size_t inputHeight = input[2];
+    size_t inputWidth = input[3];
+    size_t filterHeight = getFilterHeight(filter);
+    size_t filterWidth = getFilterWidth(filter);
+    size_t outputChannels = output[1];
+    size_t outputHeight = output[2];
+    size_t outputWidth = output[3];
+
+    real* inputData = inputs[0].data<real>();
+    real* filterData = inputs[1].data<real>();
+    real* outputData = outputs[0].data<real>();
+    bool needIm2col = isNeedIm2col(filter);
+
+    TensorShape imShape =
+        TensorShape({inputChannels / groups_, inputHeight, inputWidth});
+
+    TensorShape colShape;
+    real* colData = NULL;
+
+    size_t colHeight = inputChannels / groups_ * filterHeight * filterWidth;
+    size_t colWidth = outputHeight * outputWidth;
+    // Max col matrix height 256, Max col matrix width 1024
+    size_t stepColHeight = std::min(colHeight, static_cast<size_t>(256));
+    size_t stepColWidth = std::min(colWidth, static_cast<size_t>(2048));
+
+    if (needIm2col) {
+      colShape = TensorShape({inputChannels / groups_,
+                              filterHeight,
+                              filterWidth,
+                              outputHeight,
+                              outputWidth});
+
+      resizeBuffer<Device>(stepColHeight * stepColWidth * sizeof(real));
+      colData = reinterpret_cast<real*>(memory_->getBuf());
+    }
+
+    Im2ColMobileFunctor<real> im2col;
+    size_t inputOffset = imShape.getElements();
+    size_t outputOffset =
+        (outputChannels / groups_) * outputHeight * outputWidth;
+    size_t filterOffset = filter.getElements() / groups_;
+
+    int nStride = colWidth;
+    int kStride = colHeight;
+    for (size_t i = 0; i < batchSize; i++) {
+      for (size_t g = 0; g < groups_; g++) {
+        if (needIm2col) {
+          real beta_ = beta;
+          for (size_t colHeightStart = 0; colHeightStart < colHeight;
+               colHeightStart += stepColHeight) {
+            for (size_t colWidthStart = 0; colWidthStart < colWidth;
+                 colWidthStart += stepColWidth) {
+              int N = std::min(colWidth - colWidthStart, stepColWidth);
+              int K = std::min(colHeight - colHeightStart, stepColHeight);
+              // im2col
+              im2col(inputData + g * inputOffset,
+                     imShape,
+                     colData,
+                     colShape,
+                     strideH(),
+                     strideW(),
+                     paddingH(),
+                     paddingW(),
+                     dilationH(),
+                     dilationW(),
+                     colHeightStart,
+                     K,
+                     colWidthStart,
+                     N);
+
+              // gemm
+              int M = outputChannels / groups_;
+              BlasGemm<Device, real>::compute(
+                  false,
+                  false,
+                  M,
+                  N,
+                  K,
+                  1.0f,
+                  filterData + g * filterOffset + colHeightStart,
+                  kStride,
+                  colData,
+                  N,
+                  beta_,
+                  outputData + g * outputOffset + colWidthStart,
+                  nStride);
+            }
+            beta_ = 1.0;
+          }
+        } else {
+          int M = outputChannels / groups_;
+          int N = outputHeight * outputWidth;
+          int K = inputChannels / groups_ * filterHeight * filterWidth;
+          BlasGemm<Device, real>::compute(false,
+                                          false,
+                                          M,
+                                          N,
+                                          K,
+                                          1.0f,
+                                          filterData + g * filterOffset,
+                                          K,
+                                          inputData + g * inputOffset,
+                                          N,
+                                          beta,
+                                          outputData + g * outputOffset,
+                                          N);
+        }
+      }
+      inputData += inputChannels * inputHeight * inputWidth;
+      outputData += outputChannels * outputHeight * outputWidth;
+    }
+
+    memory_.reset();
   }
 };
 
+#endif
+
 /*
  * \brief Backward input calculation of convolution.
  */
@@ -348,7 +499,11 @@ public:
   }
 };
 
+#ifdef PADDLE_MOBILE_INFERENCE
+REGISTER_TYPED_FUNC(GemmConv, CPU, GemmConvMobileFunction);
+#else
 REGISTER_TYPED_FUNC(GemmConv, CPU, GemmConvFunction);
+#endif
 REGISTER_TYPED_FUNC(GemmConvGradInput, CPU, GemmConvGradInputFunction);
 REGISTER_TYPED_FUNC(GemmConvGradFilter, CPU, GemmConvGradFilterFunction);
 #ifdef PADDLE_WITH_CUDA

paddle/function/Im2Col.h

@@ -98,4 +98,54 @@ public:
                   int dilationWidth = 1);
 };
 
+template <class T>
+class Im2ColMobileFunctor {
+public:
+  void operator()(const T* imData,
+                  const TensorShape& imShape,
+                  T* colData,
+                  const TensorShape& colShape,
+                  int strideHeight,
+                  int strideWidth,
+                  int paddingHeight,
+                  int paddingWidth,
+                  int dilationHeight,
+                  int dilationWidth,
+                  int colHeightStart,
+                  int colHeightSize,
+                  int colWidthStart,
+                  int colWidthSize) {
+    int inputHeight = imShape[1];
+    int inputWidth = imShape[2];
+    int filterHeight = colShape[1];
+    int filterWidth = colShape[2];
+    int outputWidth = colShape[4];
+
+    for (int colh = 0; colh < colHeightSize; colh++) {
+      int wOffset = (colHeightStart + colh) % filterWidth;
+      int hOffset = ((colHeightStart + colh) / filterWidth) % filterHeight;
+      int c_im = (colHeightStart + colh) / filterWidth / filterHeight;
+
+      for (int colw = 0; colw < colWidthSize; colw++) {
+        int h = (colWidthStart + colw) / outputWidth;
+        int w = (colWidthStart + colw) % outputWidth;
+
+        int imRowIdx = h * strideHeight + hOffset * dilationHeight;
+        int imColIdx = w * strideWidth + wOffset * dilationWidth;
+        if ((imRowIdx - paddingHeight) < 0 ||
+            (imRowIdx - paddingHeight) >= inputHeight ||
+            (imColIdx - paddingWidth) < 0 ||
+            (imColIdx - paddingWidth) >= inputWidth) {
+          colData[colh * colWidthSize + colw] = static_cast<T>(0);
+        } else {
+          imRowIdx += c_im * inputHeight - paddingHeight;
+          imColIdx -= paddingWidth;
+          colData[colh * colWidthSize + colw] =
+              imData[imRowIdx * inputWidth + imColIdx];
+        }
+      }
+    }
+  }
+};
+
 }  // namespace paddle

paddle/function/Im2ColTest.cpp

@@ -138,4 +138,86 @@ TEST(Im2ColFunctor, GPU) { TestIm2ColFunctor<DEVICE_TYPE_GPU, float>(); }
 
 #endif
 
+template <class T>
+void TestIm2ColMobileFunctor() {
+  for (size_t channels : {32}) {
+    for (size_t inputHeight : {33, 100}) {
+      for (size_t inputWidth : {32, 96}) {
+        for (size_t filterHeight : {5}) {
+          for (size_t filterWidth : {7}) {
+            for (size_t stride : {2}) {
+              for (size_t padding : {1}) {
+                for (size_t dilation : {1, 3}) {
+                  size_t filterSizeH = (filterHeight - 1) * dilation + 1;
+                  size_t filterSizeW = (filterWidth - 1) * dilation + 1;
+                  if (inputHeight + 2 * padding < filterSizeH ||
+                      inputWidth + 2 * padding < filterSizeW)
+                    break;
+                  if (padding >= filterSizeH || padding >= filterSizeW) break;
+                  size_t outputHeight =
+                      (inputHeight - filterSizeH + 2 * padding) / stride + 1;
+                  size_t outputWidth =
+                      (inputWidth - filterSizeW + 2 * padding) / stride + 1;
+
+                  TensorShape imShape =
+                      TensorShape({channels, inputHeight, inputWidth});
+                  TensorShape colShape1 = TensorShape({channels,
+                                                       filterHeight,
+                                                       filterWidth,
+                                                       outputHeight,
+                                                       outputWidth});
+
+                  size_t height = channels * filterHeight * filterWidth;
+                  size_t width = outputHeight * outputWidth;
+                  VectorPtr input1 =
+                      Vector::create(imShape.getElements(), false);
+                  VectorPtr input2 =
+                      Vector::create(imShape.getElements(), false);
+                  MatrixPtr output1 =
+                      Matrix::create(height, width, false, false);
+                  MatrixPtr output2 =
+                      Matrix::create(height, width, false, false);
+                  input1->uniform(0.001, 1);
+                  input2->copyFrom(*input1);
+
+                  Im2ColFunctor<kCFO, DEVICE_TYPE_CPU, T> im2Col1;
+                  Im2ColMobileFunctor<T> im2Col2;
+                  im2Col1(input1->getData(),
+                          imShape,
+                          output1->getData(),
+                          colShape1,
+                          stride,
+                          stride,
+                          padding,
+                          padding,
+                          dilation,
+                          dilation);
+                  im2Col2(input2->getData(),
+                          imShape,
+                          output2->getData(),
+                          colShape1,
+                          stride,
+                          stride,
+                          padding,
+                          padding,
+                          dilation,
+                          dilation,
+                          0,
+                          height,
+                          0,
+                          width);
+
+                  autotest::TensorCheckEqual(*output1, *output2);
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+}
+
+TEST(Im2ColFunctor, Mobile) { TestIm2ColMobileFunctor<float>(); }
+
 }  // namespace paddle

paddle/gserver/layers/MKLDNNLRNLayer.cpp

@@ -29,7 +29,7 @@ bool MKLDNNLRNLayer::init(const LayerMap& layerMap,
   }
 
   /* the size of inputs for norm-layer is 1 */
-  CHECK_EQ(config_.inputs_size(), 1UL);
+  CHECK_EQ(config_.inputs_size(), 1);
   const NormConfig& conf = config_.inputs(0).norm_conf();
   localSize_ = conf.size();
   alpha_ = conf.scale();

paddle/memory/README.md

@@ -12,13 +12,13 @@ p = memory::Alloc(platform::CPUPlace(), 4*1024);
 To allocate 4KB memory on the 3rd GPU:
 
 ```cpp
-p = memory::Alloc(platform::GPUPlace(2), 4*1024);
+p = memory::Alloc(platform::CUDAPlace(2), 4*1024);
 ```
 
 To free memory and check the so-far used amount of memory on a place:
 
 ```cpp
-auto pl = platform::GPUPlace(0);
+auto pl = platform::CUDAPlace(0);
 p = memory::Alloc(pl, 4*1024);
 cout << memory::Used(pl);
 memory::Free(pl, p);
@@ -36,7 +36,7 @@ template <typename Place> size_t Used(Place);
 }  // namespace memory
 ```
 
-These function templates have specializations on either `platform::CPUPlace` or `platform::GPUPlace`:
+These function templates have specializations on either `platform::CPUPlace` or `platform::CUDAPlace`:
 
 ```cpp
 template<>
@@ -49,7 +49,7 @@ and
 
 ```cpp
 template<>
-void Alloc<GPUPlace>(GPUPlace p, size_t size) {
+void Alloc<CUDAPlace>(CUDAPlace p, size_t size) {
   return GetGPUBuddyAllocator(p.id)->Alloc(size);
 }
 ```
@@ -122,7 +122,7 @@ There are two implementations of `Context`:
 
 1. [`CPUContext`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context.h#L105), whose [`New` method](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context.h#L131) calls [`g_cpu_allocator.get()->New(size_t)`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context.cc#L15) to allocate the memory.
 
-1. [`CUDAContext`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context_gpu.h#L99), which has a data member [`int gpu_id_`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context_gpu.h#L202).  This looks very similar to class `majel::GPUPlace`, who also has an `int id_` data member.   `CUDAContext::New(size_t)` calls [`g_cub_allocator->DeviceAllocate(&ptr, nbytes)`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context_gpu.cu#L355) to allocate the memory.
+1. [`CUDAContext`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context_gpu.h#L99), which has a data member [`int gpu_id_`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context_gpu.h#L202).  This looks very similar to class `majel::CUDAPlace`, who also has an `int id_` data member.   `CUDAContext::New(size_t)` calls [`g_cub_allocator->DeviceAllocate(&ptr, nbytes)`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context_gpu.cu#L355) to allocate the memory.
 
 ### Majel
 

paddle/memory/detail/buddy_allocator.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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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/memory/detail/buddy_allocator.h"
 #include "glog/logging.h"

paddle/memory/detail/buddy_allocator.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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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
 

paddle/memory/detail/memory_block.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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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/memory/detail/memory_block.h"
 #include "paddle/memory/detail/meta_cache.h"

paddle/memory/detail/memory_block.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
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   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. */
+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