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

.gitignore

@@ -27,3 +27,4 @@ CMakeFiles
 cmake_install.cmake
 paddle/.timestamp
 python/paddlepaddle.egg-info/
+paddle/pybind/pybind.h

doc/design/refactorization.md

+# Design Doc: Refactorization Overview
+
+The goal of refactorizaiton include:
+
+1. Make it easy for external contributors to write new elementory computaiton operations.
+1. Make the codebase clean and readable.
+1. Introduce a new design of computation representation -- a computation graph of operators and variables.
+1. The graph representation helps implementing auto-scalable and auto fault recoverable distributed computing.
+
+## Computation Graphs
+
+1. PaddlePaddle represent the computation, training and inference of DL models, by computation graphs.
+
+  1. Please dig into [computation graphs](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/graph.md) for a solid example.
+
+1. Users write Python programs to describe the graphs and run it (locally or remotely).
+
+1. A graph is composed of *variabels* and *operators*.
+
+1. The description of graphs must be able to be serialized/deserialized, so it
+
+   1. could to be sent to the cloud for distributed execution, and
+   1. be sent to clients for mobile or enterprise deployment.
+
+1. The Python program do
+
+   1. *compilation*: runs a Python program to generate a protobuf message representation of the graph and send it to
+      1. the C++ library `libpaddle.so` for local execution,
+      1. the master process of a distributed training job for training, or
+      1. the server process of a Kubernetes serving job for distributed serving.
+   1. *execution*: according to the protobuf message, constructs instances of class `Variable` and `OperatorBase`, and run them.
+
+## Description and Realization
+
+At compile time, the Python program generates protobuf message representation of the graph, or the description of the graph.
+
+At runtime, the C++ program realizes the graph and run it.
+
+| | Representation (protobuf messages) | Realization (C++ class objects) |
+|---|---|---|
+|Data|[VarDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto#L107)|[Variable](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/variable.h#L24)|
+|Operation|[OpDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto#L35)|[Operator](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h#L64)|
+|Block|BlockDesc|Block|
+
+The word *graph* is exchangable with *block* in this document.  A graph represent computation steps and local variables as a C++/Java program block, or a pair of { and }.
+
+## Compilation and Execution
+
+1. Run an applicaton Python program to describe the graph.  In particular,
+
+   1. create VarDesc to represent local/intermediate variables,
+   1. create operators and set attributes,
+   1. validate attribute values,
+   1. inference the type and the shape of variables,
+   1. plan for memory-reuse for variables,
+   1. generate backward and optimization part of the Graph.
+   1. possiblly split the graph for distributed training.
+
+1. The invocation of `train` or `infer` in the application Python program:
+
+   1. create a new Scope instance in the [scope hierarchy](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/scope.md) for each run of a block,
+      1. realize local variables defined in the BlockDesc message in the new scope,
+      1. a scope is similar to the stack frame in programming languages,
+
+   1. create an instance of class `Block`, in which,
+      1. realize operators in the BlockDesc message,
+
+   1. run the Block by calling
+      1. `Block::Eval(vector<Variable>* targets)` for forward and backward computations, or
+      1. `Block::Eval(vector<Operator>* targets)` for optimization.
+
+
+## Intermediate Representation (IR)
+
+```text
+Compile Time -> IR -> Runtime
+```
+
+### Benefit
+
+- Optimization
+  ```text
+  Compile Time -> IR -> Optimized IR -> Runtime
+  ```
+- Send automatically partitioned IR to different nodes.
+  - Automatic data parallel
+    ```text
+    Compile Time
+    |-> Single GPU IR
+        |-> [trainer-IR-0, trainer-IR-1, pserver-IR]
+            |-> Node-0 (runs trainer-IR-0)
+            |-> Node-1 (runs trainer-IR-1)
+            |-> Node-2 (runs pserver-IR)
+    ```
+  - Automatic model parallel (planned for future)
+
+---
+
+# Operator/OpWithKernel/OpKernel
+
+![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/49caf1fb70820fb4a6c217634317c9306f361f36/op_op_with_kern_class_diagram.dot)
+
+---
+
+# Operator
+![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/dd598e8f1976f5759f58af5e5ef94738a6b2e661/op.dot)
+
+* `Operator` is the fundamental building block as the user interface.
+    * Operator stores input/output variable name, and attributes.
+    * The `InferShape` interface is used to infer output variable shapes by its input shapes.
+    * Use `Run` to compute `input variables` to `output variables`.
+
+---
+
+# OpWithKernel/Kernel
+
+![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/9d7f4eba185cf41c8e2fbfb40ae21890dbddcd39/op_with_kernel.dot)
+
+* `OpWithKernel` inherits `Operator`.
+* `OpWithKernel` contains a Kernel map.
+    * `OpWithKernel::Run` get device's kernel, and invoke `OpKernel::Compute`.
+    * `OpKernelKey` is the map key. Only device place now, but may be data type later.
+
+---
+
+# Why separate Kernel and Operator
+
+* Separate GPU and CPU code.
+    * Make Paddle can run without GPU.
+* Make one operator (which is user interface) can contain many implementations.
+    * Same mul op, different FP16, FP32 Kernel. different MKL, eigen kernel.
+---
+
+# Libraries for Kernel development
+
+* `Eigen::Tensor` contains basic math and element-wise functions.
+    * Note that `Eigen::Tensor` has broadcast implementation.
+    * Limit number of `tensor.device(dev) = ` in your code.
+* `thrust::tranform` and `std::transform`.
+    * `thrust` has the same API as C++ standard library. Using `transform` can quickly implement a customized elementwise kernel.
+    * `thrust` has more complex API, like `scan`, `reduce`, `reduce_by_key`.
+* Hand-writing `GPUKernel` and `CPU` code
+    * Do not write `.h`. CPU Kernel should be in `.cc`. CPU kernel should be in `.cu`. (`GCC` cannot compile GPU code.)
+---
+# Operator Register
+
+## Why register is necessary?
+We need a method to build mappings between Op type names and Op classes.
+
+## How to do the register?
+
+Maintain a map, whose key is the type name and value is corresponding Op constructor.
+
+---
+# The Registry Map
+
+### `OpInfoMap`
+
+`op_type(string)` -> `OpInfo`
+
+`OpInfo`:
+
+- **`creator`**: The Op constructor.
+- **`grad_op_type`**: The type of the gradient Op.
+- **`proto`**: The Op's Protobuf, including inputs, outputs and required attributes.
+- **`checker`**: Used to check attributes.
+
+---
+# Related Concepts
+
+### Op_Maker
+It's constructor takes `proto` and `checker`. They are compeleted during Op_Maker's construction. ([ScaleOpMaker](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37))
+
+### Register Macros
+```cpp
+REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type, grad_op_class)
+REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class)
+```
+
+### `USE` Macros
+make sure the registration process is executed and linked.
+
+---
+# Register Process
+1. Write Op class, as well as its gradient Op class if there is.
+2. Write Op maker class. In the constructor, describe its inputs, outputs, and attributes.
+3. Invoke macro `REGISTER_OP`. The macro will
+	1. call maker class to complete `proto` and `checker`
+	2. with the completed `proto` and `checker`, build a new key-value pair in the `OpInfoMap`
+
+4. Invoke `USE` macro in where the Op is used to make sure it is linked.
+
+---
+# Backward Module (1/2)
+### Create Backward Operator
+- Mapping from forwarding Op to backward Op
+![backward](https://gist.githubusercontent.com/dzhwinter/a6fbd4623ee76c459f7f94591fd1abf0/raw/61026ab6e518e66bde66a889bc42557a1fccff33/backward.png)
+
+---
+# Backward Module (2/2)
+### Build Backward Network
+- **Input** graph of forwarding operators
+- **Output** graph of backward operators
+- **corner case in construction**
+	- shared variable => insert `Add` operator
+	- no gradient => insert `fill_zero_grad` operator
+	- recursive netOp => call `Backward` recursively
+	- RNN Op => recursively call `Backward` on stepnet
+
+
+---
+# Scope, Variable, Tensor
+
+* `Tensor` is an n-dimension array with type.
+	* Only dims and data pointers are stored in `Tensor`.
+	* All operators on `Tensor` is written in `Operator` or global functions.
+	* variable length Tensor design [LoDTensor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/lod_tensor.md)
+* `Variable` is the inputs and outputs of an operator. Not just `Tensor`.
+	* step_scopes in RNN is a variable and not a tensor.
+* `Scope` is where variables store at.
+	* map<string/*var name */, Variable>
+	* `Scope` has a hierarchical structure. The local scope can get variable from its parent scope.
+
+---
+# Block (in design)
+## the difference with original RNNOp
+- as an operator is more intuitive than `RNNOp`,
+- offers new interface `Eval(targets)` to deduce the minimal block to `Run`,
+- fits the compile-time/ runtime separation design.
+  - during the compilation, `SymbolTable` stores `VarDesc`s and `OpDesc`s and serialize to a `BlockDesc`
+  - when graph executes, a Block with `BlockDesc` passed in creates `Op` and `Var` then `Run`
+
+---
+# Milestone
+- take Paddle/books as the main line, the requirement of the models motivates framework refactoring,
+- model migration
+  - framework development gives **priority support** to model migration, for example,
+    - the MNIST demo needs a Python interface,
+    - the RNN models require the framework to support `LoDTensor`.
+  - determine some timelines,
+  - heavily-relied Ops need to be migrated first,
+  - different models can be migrated parallelly.
+- improve the framework at the same time
+- accept imperfection, concentrated on solving the specific problem at the right price.
+
+---
+# Control the migration quality
+- compare the performance of migrated models with old ones.
+- follow google C style
+- build the automatic workflow of generating Python/C++ documentations
+  - the documentation of layers and ops should be written inside the code
+  - take the documentation quality into account when doing PR
+  - preview the documentations, read and improve them from users' perspective

paddle/cuda/include/hl_cuda_cudnn.h

@@ -22,10 +22,10 @@ limitations under the License. */
  */
 typedef enum {
   HL_POOLING_MAX = 0,
-  // average includes padded values
-  HL_POOLING_AVERAGE = 1,
   // average does not include padded values
-  HL_POOLING_AVERAGE_EXCLUDE_PADDING = 2,
+  HL_POOLING_AVERAGE = 1,
+  // average includes padded values
+  HL_POOLING_AVERAGE_INCLUDE_PADDING = 2,
   HL_POOLING_END
 } hl_pooling_mode_t;
 

paddle/cuda/src/hl_cuda_cnn.cu

@@ -211,13 +211,11 @@ __global__ void KeAvgPoolForward(const int nthreads,
 
     int hstart = ph * strideH - padH;
     int wstart = pw * strideW - padW;
-    int hend = min(hstart + sizeY, height + padH);
-    int wend = min(wstart + sizeX, width + padW);
-    int pool_size = (hend - hstart) * (wend - wstart);
+    int hend = min(hstart + sizeY, height);
+    int wend = min(wstart + sizeX, width);
     hstart = max(hstart, 0);
     wstart = max(wstart, 0);
-    hend = min(hend, height);
-    wend = min(wend, width);
+    int pool_size = (hend - hstart) * (wend - wstart);
 
     real aveval = 0;
     inputData += (frameNum * channels + c) * height * width;
@@ -299,12 +297,14 @@ __global__ void KeAvgPoolBackward(const int nthreads,
     outGrad += (frameNum * outStride + offsetC * pooledH * pooledW);
 
     for (int ph = phstart; ph < phend; ++ph) {
+      int hstart = ph * strideH - padH;
+      int hend = min(hstart + sizeY, height);
+      hstart = max(hstart, 0);
       for (int pw = pwstart; pw < pwend; ++pw) {
         // figure out the pooling size
-        int hstart = ph * strideH - padH;
         int wstart = pw * strideW - padW;
-        int hend = min(hstart + sizeY, height + padH);
-        int wend = min(wstart + sizeX, width + padW);
+        int wend = min(wstart + sizeX, width);
+        wstart = max(wstart, 0);
         int poolsize = (hend - hstart) * (wend - wstart);
         gradient += outGrad[ph * pooledW + pw] / poolsize;
       }
@@ -600,16 +600,13 @@ __global__ void KeAvgPool3DForward(const int nthreads,
     int dstart = pd * strideD - padD;
     int hstart = ph * strideH - padH;
     int wstart = pw * strideW - padW;
-    int dend = min(dstart + sizeZ, depth + padD);
-    int hend = min(hstart + sizeY, height + padH);
-    int wend = min(wstart + sizeX, width + padW);
-    int pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart);
+    int dend = min(dstart + sizeZ, depth);
+    int hend = min(hstart + sizeY, height);
+    int wend = min(wstart + sizeX, width);
     dstart = max(dstart, 0);
     hstart = max(hstart, 0);
     wstart = max(wstart, 0);
-    dend = min(dend, depth);
-    hend = min(hend, height);
-    wend = min(wend, width);
+    int pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart);
 
     real aveval = 0;
     inputData += (frameNum * channels + c) * depth * height * width;
@@ -712,15 +709,18 @@ __global__ void KeAvgPool3DBackward(const int nthreads,
     outGrad += (frameNum * channels + offsetC) * pooledD * pooledH * pooledW;
 
     for (int pd = pdstart; pd < pdend; ++pd) {
+      int dstart = pd * strideD - padD;
+      int dend = min(dstart + sizeZ, depth);
+      dstart = max(dstart, 0);
       for (int ph = phstart; ph < phend; ++ph) {
+        int hstart = ph * strideH - padH;
+        int hend = min(hstart + sizeY, height);
+        hstart = max(hstart, 0);
         for (int pw = pwstart; pw < pwend; ++pw) {
           // figure out the pooling size
-          int dstart = pd * strideD - padD;
-          int hstart = ph * strideH - padH;
           int wstart = pw * strideW - padW;
-          int dend = min(dstart + sizeZ, depth + padD);
-          int hend = min(hstart + sizeY, height + padH);
-          int wend = min(wstart + sizeX, width + padW);
+          int wend = min(wstart + sizeX, width);
+          wstart = max(wstart, 0);
           int poolsize = (dend - dstart) * (hend - hstart) * (wend - wstart);
           gradient += outGrad[(pd * pooledH + ph) * pooledW + pw] / poolsize;
         }

paddle/cuda/src/hl_cuda_cudnn.cc

@@ -432,11 +432,11 @@ void hl_create_pooling_descriptor(hl_pooling_descriptor* pooling_desc,
       cudnn_mode = CUDNN_POOLING_MAX;
       break;
     case HL_POOLING_AVERAGE:
-      cudnn_mode = CUDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING;
-      break;
-    case HL_POOLING_AVERAGE_EXCLUDE_PADDING:
       cudnn_mode = CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING;
       break;
+    case HL_POOLING_AVERAGE_INCLUDE_PADDING:
+      cudnn_mode = CUDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING;
+      break;
     default:
       LOG(FATAL) << "parameter mode error";
   }

paddle/framework/operator.cc

@@ -22,14 +22,14 @@ namespace framework {
 template <>
 Eigen::DefaultDevice& ExecutionContext::GetEigenDevice<
     platform::CPUPlace, Eigen::DefaultDevice>() const {
-  return *device_context_->get_eigen_device<Eigen::DefaultDevice>();
+  return *device_context_.get_eigen_device<Eigen::DefaultDevice>();
 }
 
 #ifndef PADDLE_ONLY_CPU
 template <>
 Eigen::GpuDevice&
 ExecutionContext::GetEigenDevice<platform::GPUPlace, Eigen::GpuDevice>() const {
-  return *device_context_->get_eigen_device<Eigen::GpuDevice>();
+  return *device_context_.get_eigen_device<Eigen::GpuDevice>();
 }
 #endif
 

paddle/framework/operator.h

@@ -366,7 +366,7 @@ struct EigenDeviceConverter<platform::GPUPlace> {
 class ExecutionContext : public InferShapeContext {
  public:
   ExecutionContext(const OperatorBase& op, const Scope& scope,
-                   const platform::DeviceContext* device_context)
+                   const platform::DeviceContext& device_context)
       : InferShapeContext(op, scope), device_context_(device_context) {}
 
   template <typename PlaceType,
@@ -374,9 +374,9 @@ class ExecutionContext : public InferShapeContext {
                 typename EigenDeviceConverter<PlaceType>::EigenDeviceType>
   DeviceType& GetEigenDevice() const;
 
-  platform::Place GetPlace() const { return device_context_->GetPlace(); }
+  platform::Place GetPlace() const { return device_context_.GetPlace(); }
 
-  const platform::DeviceContext* device_context() const {
+  const platform::DeviceContext& device_context() const {
     return device_context_;
   }
 
@@ -401,7 +401,8 @@ class ExecutionContext : public InferShapeContext {
     return res;
   }
 
-  const platform::DeviceContext* device_context_;
+ private:
+  const platform::DeviceContext& device_context_;
 };
 
 template <>
@@ -461,7 +462,7 @@ class OperatorWithKernel : public OperatorBase {
   void Run(const Scope& scope,
            const platform::DeviceContext& dev_ctx) const final {
     auto& opKernel = AllOpKernels().at(type_).at(OpKernelKey(dev_ctx));
-    opKernel->Compute(ExecutionContext(*this, scope, &dev_ctx));
+    opKernel->Compute(ExecutionContext(*this, scope, dev_ctx));
   }
 
   static std::unordered_map<std::string /* op_type */, OpKernelMap>&

paddle/gserver/layers/CudnnPoolLayer.cpp

@@ -29,9 +29,9 @@ bool CudnnPoolLayer::typeCheck(const std::string &poolType,
     if (mode) {
       *mode = HL_POOLING_AVERAGE;
     }
-  } else if (poolType == "cudnn-avg-excl-pad-pool") {
+  } else if (poolType == "cudnn-avg-incl-pad-pool") {
     if (mode) {
-      *mode = HL_POOLING_AVERAGE_EXCLUDE_PADDING;
+      *mode = HL_POOLING_AVERAGE_INCLUDE_PADDING;
     }
   } else {
     return false;

paddle/gserver/layers/MKLDNNPoolLayer.cpp

+/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "MKLDNNPoolLayer.h"
+#include "paddle/math/MathUtils.h"
+#include "paddle/utils/Logging.h"
+
+using namespace mkldnn;  // NOLINT
+typedef memory::format format;
+
+namespace paddle {
+
+REGISTER_LAYER(mkldnn_pool, MKLDNNPoolLayer);
+
+bool MKLDNNPoolLayer::init(const LayerMap& layerMap,
+                           const ParameterMap& parameterMap) {
+  if (!MKLDNNLayer::init(layerMap, parameterMap)) {
+    return false;
+  }
+
+  /* the size of inputs for pool-layer is 1 */
+  CHECK_EQ(config_.inputs_size(), 1);
+  const PoolConfig& conf = config_.inputs(0).pool_conf();
+  ic_ = conf.channels();
+  ih_ = conf.img_size_y();
+  iw_ = conf.img_size();
+  oc_ = ic_;
+  oh_ = conf.output_y();
+  ow_ = conf.output_x();
+  fh_ = conf.size_y();
+  fw_ = conf.size_x();
+  ph_ = conf.padding_y();
+  pw_ = conf.padding();
+  sh_ = conf.stride_y();
+  sw_ = conf.stride();
+
+  const std::string& type = conf.pool_type();
+  if (type == "max-projection") {
+    poolAlgo_ = algorithm::pooling_max;
+  } else if (type == "avg-projection") {
+    // paddle only use exclude_padding
+    poolAlgo_ = algorithm::pooling_avg_exclude_padding;
+  } else {
+    LOG(FATAL) << "unknow pooling type!";
+  }
+  return true;
+}
+
+void MKLDNNPoolLayer::reshape(
+    int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) {
+  reshapeInput(bs, ih, iw);
+  // ic_ and oc can not be changed
+  CHECK_EQ(inputElemenCnt_ / bs / ih / iw, (size_t)ic)
+      << "Input channel can not be changed";
+
+  // cal output sizes
+  // paddle used false caffeMode for pooling
+  oh = outputSize(ih, fh_, ph_, sh_, false);
+  ow = outputSize(iw, fw_, pw_, sw_, false);
+  reshapeOutput(oh, ow);
+
+  resizeOutput(bs, oc * oh * ow);
+
+  printSizeInfo();
+}
+
+void MKLDNNPoolLayer::resetFwd(std::vector<primitive>& pipeline,
+                               MKLDNNMatrixPtr& in,
+                               MKLDNNMatrixPtr& wgt,
+                               MKLDNNMatrixPtr& bias,
+                               MKLDNNMatrixPtr& out) {
+  resetFwdBuffers(in, out);
+
+  resetFwdPD(fwdPD_, in, out);
+
+  resetFwdPipeline(pipeline, fwdPD_, in, out);
+
+  printValueFormatFlow();
+}
+
+void MKLDNNPoolLayer::resetBwd(std::vector<primitive>& pipeline,
+                               MKLDNNMatrixPtr& in,
+                               MKLDNNMatrixPtr& wgt,
+                               MKLDNNMatrixPtr& bias,
+                               MKLDNNMatrixPtr& out) {
+  std::shared_ptr<pool_bwd::primitive_desc> pd;
+
+  resetBwdBuffers(in, out);
+
+  resetBwdPD(pd, in, out);
+
+  resetBwdPipeline(pipeline, pd, in, out);
+
+  printGradFormatFlow();
+}
+
+void MKLDNNPoolLayer::updateInputData() {
+  inVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
+}
+
+void MKLDNNPoolLayer::resetFwdBuffers(MKLDNNMatrixPtr& in,
+                                      MKLDNNMatrixPtr& out) {
+  resetInValue(in);
+
+  resetOutValue(out);
+}
+
+void MKLDNNPoolLayer::resetInValue(MKLDNNMatrixPtr& in) {
+  if (inputIsOnlyMKLDNN()) {
+    const MatrixPtr& dnnIn = getInputValue(0);
+    in = std::dynamic_pointer_cast<MKLDNNMatrix>(dnnIn);
+    CHECK(in) << "Input should be MKLDNNMatrix";
+  } else {
+    CHECK_EQ(getPrev(0)->getDeviceId(), CPU_DEVICE) << "Only support CPU yet";
+    const MatrixPtr& cpuIn = getInputValue(0, CPU_DEVICE);
+    in = MKLDNNMatrix::create(
+        cpuIn, {bs_, ic_, ih_, iw_}, format::nchw, engine_);
+  }
+}
+
+void MKLDNNPoolLayer::resetOutValue(MKLDNNMatrixPtr& out) {
+  CHECK(inVal_) << "Should reset input value first";
+  memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
+  out = MKLDNNMatrix::create(
+      output_.value, outDims, inVal_->getFormat(), engine_);
+  output_.value = std::dynamic_pointer_cast<Matrix>(out);
+
+  // create reorder if output value has cpu device and pd do not match
+  cpuOutVal_ = nullptr;
+  cvtOutVal_ = nullptr;
+  if (!outputIsOnlyMKLDNN()) {
+    const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).value;
+    cpuOutVal_ = MKLDNNMatrix::create(cpuOut, outDims, format::nchw, engine_);
+    if (cpuOutVal_->getPrimitiveDesc() != out->getPrimitiveDesc()) {
+      cvtOutVal_ = MKLDNNMatrix::createReorder(out, cpuOutVal_);
+      CHECK(cvtOutVal_) << "should not be emptry";
+    } else {
+      // CPU output share the same data of MKLDNN output
+      cpuOut->setData(out->getData());
+      cpuOutVal_ = out;
+    }
+  }
+}
+
+void MKLDNNPoolLayer::resetFwdPD(std::shared_ptr<pool_fwd::primitive_desc>& pd,
+                                 MKLDNNMatrixPtr in,
+                                 MKLDNNMatrixPtr out) {
+  memory::dims inDims = memory::dims{bs_, ic_, ih_, iw_};
+  memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
+  memory::dims kernels = memory::dims{fh_, fw_};
+  memory::dims strides = memory::dims{sh_, sw_};
+  memory::dims padL = memory::dims{ph_, pw_};
+  memory::dims padR = getPaddingR();
+  padding_kind padKind = padding_kind::zero;
+  prop_kind pk = passType_ == PASS_TEST ? prop_kind::forward_scoring
+                                        : prop_kind::forward_training;
+  auto fwdDesc = pool_fwd::desc(pk,
+                                poolAlgo_,
+                                in->getMemoryDesc(),
+                                out->getMemoryDesc(),
+                                strides,
+                                kernels,
+                                padL,
+                                padR,
+                                padKind);
+  pd.reset(new pool_fwd::primitive_desc(fwdDesc, engine_));
+
+  // prepare workspace if necessary
+  workspace_ =
+      (passType_ != PASS_TEST && poolAlgo_ == algorithm::pooling_max)
+          ? std::make_shared<memory>(memory(pd->workspace_primitive_desc()))
+          : nullptr;
+}
+
+void MKLDNNPoolLayer::resetFwdPipeline(
+    std::vector<primitive>& pipeline,
+    std::shared_ptr<pool_fwd::primitive_desc>& pd,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& out) {
+  pipeline.clear();
+  fwd_ = workspace_
+             ? std::make_shared<pool_fwd>(pool_fwd(*pd, *in, *out, *workspace_))
+             : std::make_shared<pool_fwd>(pool_fwd(*pd, *in, *out));
+  pipeline.push_back(*fwd_);
+
+  if (cvtOutVal_) {
+    pipeline.push_back(*cvtOutVal_);
+  }
+}
+
+void MKLDNNPoolLayer::resetBwdBuffers(MKLDNNMatrixPtr& in,
+                                      MKLDNNMatrixPtr& out) {
+  resetOutGrad(out);
+
+  resetInGrad(in);
+}
+void MKLDNNPoolLayer::resetOutGrad(MKLDNNMatrixPtr& out) {
+  CHECK(outVal_) << "Should have output value";
+  out = MKLDNNMatrix::create(output_.grad, outVal_->getPrimitiveDesc());
+
+  // create reorder if output value has cpu device and pd do not match
+  cpuOutGrad_ = nullptr;
+  cvtOutGrad_ = nullptr;
+  if (!outputIsOnlyMKLDNN()) {
+    const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).grad;
+    cpuOutGrad_ = MKLDNNMatrix::create(
+        cpuOut, memory::dims{bs_, oc_, oh_, ow_}, format::nchw, engine_);
+    if (cpuOutGrad_->getPrimitiveDesc() != out->getPrimitiveDesc()) {
+      cvtOutGrad_ = MKLDNNMatrix::createReorder(cpuOutGrad_, out);
+      CHECK(cvtOutGrad_) << "should not be emptry";
+    } else {
+      // share the same data of CPU output
+      output_.grad->setData(cpuOut->getData());
+      out = cpuOutGrad_;
+    }
+  }
+}
+
+void MKLDNNPoolLayer::resetInGrad(MKLDNNMatrixPtr& in) {
+  in = nullptr;
+  const MatrixPtr& inGrad = inputLayers_[0]->getOutput().grad;
+  if (inGrad == nullptr) {
+    return;
+  }
+  CHECK(inVal_);
+  in = MKLDNNMatrix::create(inGrad, inVal_->getPrimitiveDesc());
+}
+
+void MKLDNNPoolLayer::resetBwdPD(std::shared_ptr<pool_bwd::primitive_desc>& pd,
+                                 MKLDNNMatrixPtr& in,
+                                 MKLDNNMatrixPtr& out) {
+  memory::dims kernels = memory::dims{fh_, fw_};
+  memory::dims strides = memory::dims{sh_, sw_};
+  memory::dims padL = memory::dims{ph_, pw_};
+  memory::dims padR = getPaddingR();
+  CHECK(in);
+  CHECK(out);
+  auto bwdDesc = pool_bwd::desc(poolAlgo_,
+                                in->getMemoryDesc(),
+                                out->getMemoryDesc(),
+                                strides,
+                                kernels,
+                                padL,
+                                padR,
+                                padding_kind::zero);
+  pd.reset(new pool_bwd::primitive_desc(bwdDesc, engine_, *fwdPD_));
+}
+
+void MKLDNNPoolLayer::resetBwdPipeline(
+    std::vector<primitive>& pipeline,
+    std::shared_ptr<pool_bwd::primitive_desc>& pd,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& out) {
+  pipeline.clear();
+  if (cvtOutGrad_) {
+    pipeline.push_back(*cvtOutGrad_);
+  }
+
+  bwdData_ =
+      workspace_
+          ? std::make_shared<pool_bwd>(pool_bwd(*pd, *out, *workspace_, *in))
+          : std::make_shared<pool_bwd>(pool_bwd(*pd, *out, *in));
+  pipeline.push_back(*bwdData_);
+}
+
+}  // namespace paddle

paddle/gserver/layers/MKLDNNPoolLayer.h

+/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include "MKLDNNLayer.h"
+#include "mkldnn.hpp"
+
+namespace paddle {
+typedef mkldnn::pooling_forward pool_fwd;
+typedef mkldnn::pooling_backward pool_bwd;
+
+/**
+ * @brief A subclass of MKLDNNLayer pool layer.
+ *
+ * The config file api is mkldnn_pool
+ */
+class MKLDNNPoolLayer : public MKLDNNLayer {
+protected:
+  // padding height and width
+  int ph_, pw_;
+  // stride height and width
+  int sh_, sw_;
+  // filter(kenerl) height and width
+  int fh_, fw_;
+
+  // pooling_avg or pooling_max
+  mkldnn::algorithm poolAlgo_;
+
+  // MKLDNNMatrixPtr which should be created from CPU Device
+  MKLDNNMatrixPtr cpuOutVal_;
+  MKLDNNMatrixPtr cpuOutGrad_;
+  // convert handle between CPU device and MKLDNN device
+  std::shared_ptr<mkldnn::reorder> cvtOutVal_;
+  std::shared_ptr<mkldnn::reorder> cvtOutGrad_;
+
+  // save forward primitive_desc, which can be used backward
+  std::shared_ptr<pool_fwd::primitive_desc> fwdPD_;
+  // according to https://github.com/01org/mkl-dnn/blob/master/tests/gtests/
+  // test_pooling_forward.cpp, pool need workspace for backward
+  std::shared_ptr<mkldnn::memory> workspace_;
+
+public:
+  explicit MKLDNNPoolLayer(const LayerConfig& config) : MKLDNNLayer(config) {}
+
+  ~MKLDNNPoolLayer() {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void reshape(
+      int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) override;
+
+  void resetFwd(std::vector<mkldnn::primitive>& pipeline,
+                MKLDNNMatrixPtr& in,
+                MKLDNNMatrixPtr& wgt,
+                MKLDNNMatrixPtr& bias,
+                MKLDNNMatrixPtr& out) override;
+
+  void resetBwd(std::vector<mkldnn::primitive>& pipeline,
+                MKLDNNMatrixPtr& in,
+                MKLDNNMatrixPtr& wgt,
+                MKLDNNMatrixPtr& bias,
+                MKLDNNMatrixPtr& out) override;
+
+  void updateInputData() override;
+
+  void printSizeInfo() override {
+    MKLDNNLayer::printSizeInfo();
+    VLOG(MKLDNN_SIZES) << getName() << ": fh: " << fh_ << ", fw: " << fw_
+                       << ": ph: " << ph_ << ", pw: " << pw_ << ", sh: " << sh_
+                       << ", sw: " << sw_;
+  }
+
+protected:
+  /**
+   * Forward functions: reset buffers(input, output),
+   *                    reset primitive descriptor,
+   *                    reset pipeline.
+   */
+  void resetFwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& out);
+  void resetInValue(MKLDNNMatrixPtr& in);
+  void resetOutValue(MKLDNNMatrixPtr& out);
+  void resetFwdPD(std::shared_ptr<pool_fwd::primitive_desc>& pd,
+                  MKLDNNMatrixPtr in,
+                  MKLDNNMatrixPtr out);
+  void resetFwdPipeline(std::vector<mkldnn::primitive>& pipeline,
+                        std::shared_ptr<pool_fwd::primitive_desc>& pd,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& out);
+
+  /**
+   * Backward functions: reset buffers(input, output),
+   *                     reset primitive descriptor,
+   *                     reset pipeline.
+   */
+  void resetBwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& out);
+  void resetOutGrad(MKLDNNMatrixPtr& out);
+  void resetInGrad(MKLDNNMatrixPtr& in);
+  void resetBwdPD(std::shared_ptr<pool_bwd::primitive_desc>& pd,
+                  MKLDNNMatrixPtr& in,
+                  MKLDNNMatrixPtr& out);
+  void resetBwdPipeline(std::vector<mkldnn::primitive>& pipeline,
+                        std::shared_ptr<pool_bwd::primitive_desc>& pd,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& out);
+
+  /**
+   * get padding_r according to
+   * https://github.com/01org/mkl-dnn/blob/master/tests/gtests/
+   * test_pooling_forward.cpp
+   */
+  mkldnn::memory::dims getPaddingR() const {
+    mkldnn::memory::dims padR = {ph_, pw_};
+    for (int i = 0; i < 2; ++i) {
+      if ((ih_ + ph_ + padR[0] - fh_) / sh_ + 1 < oh_) {
+        ++padR[0];
+      }
+      if ((iw_ + pw_ + padR[1] - fw_) / sw_ + 1 < ow_) {
+        ++padR[1];
+      }
+    }
+    return padR;
+  }
+};
+
+}  // namespace paddle

paddle/gserver/tests/test_MKLDNN.cpp

@@ -141,6 +141,68 @@ TEST(MKLDNNLayer, ConvLayer) {
   testConvLayer({4, 4, 16, 3, 3, 16, 3, 3, 3, 3, 1, 1, 1, 1, 1, 1});
 }
 
+struct testPoolDesc {
+  int bs, ch;  // input channel and output channel are the same
+  int ih, iw;
+  int oh, ow;
+  int fh, fw;
+  int ph, pw;
+  int sh, sw;
+};
+
+void testPoolLayer(const testPoolDesc& pm) {
+  const std::string compareTypes[] = {"mkldnn_pool", "pool"};
+  TestConfig cfg;
+  cfg.layerConfig.set_type(compareTypes[0]);
+  cfg.layerConfig.set_size(pm.ch * pm.oh * pm.ow);
+  cfg.inputDefs.push_back(
+      {INPUT_DATA,
+       "layer_0",
+       /* size of input layer= */ size_t(pm.ch * pm.ih * pm.iw),
+       0});
+  LayerInputConfig* input = cfg.layerConfig.add_inputs();
+  PoolConfig* pool = input->mutable_pool_conf();
+  // pool->set_pool_type(poolType);
+  pool->set_channels(pm.ch);
+  pool->set_img_size(pm.iw);
+  pool->set_img_size_y(pm.ih);
+  pool->set_output_x(pm.ow);
+  pool->set_output_y(pm.oh);
+  pool->set_size_x(pm.fw);
+  pool->set_size_y(pm.fh);
+  pool->set_padding(pm.pw);
+  pool->set_padding_y(pm.ph);
+  pool->set_stride(pm.sw);
+  pool->set_stride_y(pm.sh);
+
+  int oh = outputSize(pm.ih, pm.fh, pm.ph, pm.sh, false);
+  int ow = outputSize(pm.iw, pm.fw, pm.pw, pm.sw, false);
+  CHECK_EQ(ow, pm.ow) << "output size check failed";
+  CHECK_EQ(oh, pm.oh) << "output size check failed";
+
+  MKLDNNTester tester;
+  for (auto type : {"max-projection", "avg-projection"}) {
+    pool->set_pool_type(type);
+    TestConfig ref = cfg;
+    ref.layerConfig.set_type(compareTypes[1]);
+    for (auto bs : {pm.bs, 1}) {
+      tester.run(cfg, ref, bs, pm.ih, pm.iw);
+    }
+  }
+}
+
+TEST(MkldnnLayer, PoolLayer) {
+  /* bs, ch, ih, iw, oh, ow, fh, fw, ph, pw, sh, sw*/
+  testPoolLayer({2, 1, 4, 4, 2, 2, 3, 3, 0, 0, 2, 2});
+  testPoolLayer({10, 8, 16, 16, 8, 8, 2, 2, 0, 0, 2, 2});
+  testPoolLayer({4, 2, 5, 5, 3, 3, 3, 3, 1, 1, 2, 2});
+  testPoolLayer({8, 16, 56, 56, 28, 28, 3, 3, 0, 0, 2, 2});
+  testPoolLayer({8, 16, 14, 14, 7, 7, 3, 3, 0, 0, 2, 2});
+  testPoolLayer({4, 16, 7, 7, 1, 1, 7, 7, 0, 0, 1, 1});
+  testPoolLayer({4, 2, 5, 5, 3, 3, 5, 5, 1, 1, 1, 1});
+  testPoolLayer({2, 8, 56, 56, 29, 29, 3, 3, 1, 1, 2, 2});
+}
+
 // TODO(TJ): add branch test
 
 int main(int argc, char** argv) {

paddle/math/tests/test_matrixCompare.cpp

@@ -825,9 +825,8 @@ void testMaxPoolFwdBwd(int numSamples,
                        int strideW,
                        int padH,
                        int padW) {
-  int outH = 0, outW = 0;
-  outH = (imgSizeH - ksizeH + 2 * padH + strideH - 1) / strideH + 1;
-  outW = (imgSizeW - ksizeW + 2 * padW + strideW - 1) / strideW + 1;
+  int outH = outputSize(imgSizeH, ksizeH, padH, strideH, true);
+  int outW = outputSize(imgSizeW, ksizeW, padW, strideW, true);
 
   int inWidth = imgSizeH * imgSizeW * channels;
   MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false);
@@ -927,9 +926,8 @@ void testAvgPoolFwdBwd(int numSamples,
                        int strideW,
                        int padH,
                        int padW) {
-  int outH = 0, outW = 0;
-  outH = (imgSizeH - ksizeH + 2 * padH + strideH - 1) / strideH + 1;
-  outW = (imgSizeW - ksizeW + 2 * padW + strideW - 1) / strideW + 1;
+  int outH = outputSize(imgSizeH, ksizeH, padH, strideH, true);
+  int outW = outputSize(imgSizeW, ksizeW, padW, strideW, true);
 
   int inWidth = imgSizeH * imgSizeW * channels;
   MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false);

paddle/operators/math/math_function.cc

@@ -19,12 +19,13 @@ namespace operators {
 namespace math {
 
 template <>
-void gemm<platform::CPUPlace, float>(const CBLAS_TRANSPOSE transA,
+void gemm<platform::CPUPlace, float>(const platform::DeviceContext& context,
+                                     const CBLAS_TRANSPOSE transA,
                                      const CBLAS_TRANSPOSE transB, const int M,
                                      const int N, const int K,
                                      const float alpha, const float* A,
-                                     const float* B, const float beta, float* C,
-                                     platform::DeviceContext* context) {
+                                     const float* B, const float beta,
+                                     float* C) {
   int lda = (transA == CblasNoTrans) ? K : M;
   int ldb = (transB == CblasNoTrans) ? N : K;
   int ldc = N;
@@ -33,13 +34,13 @@ void gemm<platform::CPUPlace, float>(const CBLAS_TRANSPOSE transA,
 }
 
 template <>
-void gemm<platform::CPUPlace, double>(const CBLAS_TRANSPOSE transA,
+void gemm<platform::CPUPlace, double>(const platform::DeviceContext& context,
+                                      const CBLAS_TRANSPOSE transA,
                                       const CBLAS_TRANSPOSE transB, const int M,
                                       const int N, const int K,
                                       const double alpha, const double* A,
                                       const double* B, const double beta,
-                                      double* C,
-                                      platform::DeviceContext* context) {
+                                      double* C) {
   int lda = (transA == CblasNoTrans) ? K : M;
   int ldb = (transB == CblasNoTrans) ? N : K;
   int ldc = N;
@@ -48,13 +49,10 @@ void gemm<platform::CPUPlace, double>(const CBLAS_TRANSPOSE transA,
 }
 
 template <>
-void matmul<platform::CPUPlace, float>(const framework::Tensor& matrix_a,
-                                       bool trans_a,
-                                       const framework::Tensor& matrix_b,
-                                       bool trans_b, float alpha,
-                                       framework::Tensor* matrix_out,
-                                       float beta,
-                                       platform::DeviceContext* context) {
+void matmul<platform::CPUPlace, float>(
+    const platform::DeviceContext& context, const framework::Tensor& matrix_a,
+    bool trans_a, const framework::Tensor& matrix_b, bool trans_b, float alpha,
+    framework::Tensor* matrix_out, float beta) {
   auto dim_a = matrix_a.dims();
   auto dim_b = matrix_b.dims();
   auto dim_out = matrix_out->dims();
@@ -74,18 +72,15 @@ void matmul<platform::CPUPlace, float>(const framework::Tensor& matrix_a,
   CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans;
 
   gemm<platform::CPUPlace, float>(
-      transA, transB, M, N, K, alpha, matrix_a.data<float>(),
-      matrix_b.data<float>(), beta, matrix_out->data<float>(), context);
+      context, transA, transB, M, N, K, alpha, matrix_a.data<float>(),
+      matrix_b.data<float>(), beta, matrix_out->data<float>());
 }
 
 template <>
-void matmul<platform::CPUPlace, double>(const framework::Tensor& matrix_a,
-                                        bool trans_a,
-                                        const framework::Tensor& matrix_b,
-                                        bool trans_b, double alpha,
-                                        framework::Tensor* matrix_out,
-                                        double beta,
-                                        platform::DeviceContext* context) {
+void matmul<platform::CPUPlace, double>(
+    const platform::DeviceContext& context, const framework::Tensor& matrix_a,
+    bool trans_a, const framework::Tensor& matrix_b, bool trans_b, double alpha,
+    framework::Tensor* matrix_out, double beta) {
   auto dim_a = matrix_a.dims();
   auto dim_b = matrix_b.dims();
   auto dim_out = matrix_out->dims();
@@ -105,8 +100,8 @@ void matmul<platform::CPUPlace, double>(const framework::Tensor& matrix_a,
   CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans;
 
   gemm<platform::CPUPlace, double>(
-      transA, transB, M, N, K, alpha, matrix_a.data<double>(),
-      matrix_b.data<double>(), beta, matrix_out->data<double>(), context);
+      context, transA, transB, M, N, K, alpha, matrix_a.data<double>(),
+      matrix_b.data<double>(), beta, matrix_out->data<double>());
 }
 
 }  // namespace math

paddle/operators/math/math_function.cu

@@ -19,12 +19,13 @@ namespace operators {
 namespace math {
 
 template <>
-void gemm<platform::GPUPlace, float>(const CBLAS_TRANSPOSE transA,
+void gemm<platform::GPUPlace, float>(const platform::DeviceContext& context,
+                                     const CBLAS_TRANSPOSE transA,
                                      const CBLAS_TRANSPOSE transB, const int M,
                                      const int N, const int K,
                                      const float alpha, const float* A,
-                                     const float* B, const float beta, float* C,
-                                     platform::DeviceContext* context) {
+                                     const float* B, const float beta,
+                                     float* C) {
   // Note that cublas follows fortran order, so the order is different from
   // the cblas convention.
   int lda = (transA == CblasNoTrans) ? K : M;
@@ -35,18 +36,19 @@ void gemm<platform::GPUPlace, float>(const CBLAS_TRANSPOSE transA,
       (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
 
   PADDLE_ENFORCE(platform::dynload::cublasSgemm(
-      reinterpret_cast<platform::CUDADeviceContext*>(context)->cublas_handle(),
+      reinterpret_cast<const platform::CUDADeviceContext&>(context)
+          .cublas_handle(),
       cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, lda, &beta, C, N));
 }
 
 template <>
-void gemm<platform::GPUPlace, double>(const CBLAS_TRANSPOSE transA,
+void gemm<platform::GPUPlace, double>(const platform::DeviceContext& context,
+                                      const CBLAS_TRANSPOSE transA,
                                       const CBLAS_TRANSPOSE transB, const int M,
                                       const int N, const int K,
                                       const double alpha, const double* A,
                                       const double* B, const double beta,
-                                      double* C,
-                                      platform::DeviceContext* context) {
+                                      double* C) {
   // Note that cublas follows fortran order, so the order is different from
   // the cblas convention.
   int lda = (transA == CblasNoTrans) ? K : M;
@@ -56,18 +58,16 @@ void gemm<platform::GPUPlace, double>(const CBLAS_TRANSPOSE transA,
   cublasOperation_t cuTransB =
       (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
   PADDLE_ENFORCE(platform::dynload::cublasDgemm(
-      reinterpret_cast<platform::CUDADeviceContext*>(context)->cublas_handle(),
+      reinterpret_cast<const platform::CUDADeviceContext&>(context)
+          .cublas_handle(),
       cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, lda, &beta, C, N));
 }
 
 template <>
-void matmul<platform::GPUPlace, float>(const framework::Tensor& matrix_a,
-                                       bool trans_a,
-                                       const framework::Tensor& matrix_b,
-                                       bool trans_b, float alpha,
-                                       framework::Tensor* matrix_out,
-                                       float beta,
-                                       platform::DeviceContext* context) {
+void matmul<platform::GPUPlace, float>(
+    const platform::DeviceContext& context, const framework::Tensor& matrix_a,
+    bool trans_a, const framework::Tensor& matrix_b, bool trans_b, float alpha,
+    framework::Tensor* matrix_out, float beta) {
   auto dim_a = matrix_a.dims();
   auto dim_b = matrix_b.dims();
   auto dim_out = matrix_out->dims();
@@ -87,18 +87,15 @@ void matmul<platform::GPUPlace, float>(const framework::Tensor& matrix_a,
   CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans;
 
   gemm<platform::GPUPlace, float>(
-      transA, transB, M, N, K, alpha, matrix_a.data<float>(),
-      matrix_b.data<float>(), beta, matrix_out->data<float>(), context);
+      context, transA, transB, M, N, K, alpha, matrix_a.data<float>(),
+      matrix_b.data<float>(), beta, matrix_out->data<float>());
 }
 
 template <>
-void matmul<platform::GPUPlace, double>(const framework::Tensor& matrix_a,
-                                        bool trans_a,
-                                        const framework::Tensor& matrix_b,
-                                        bool trans_b, double alpha,
-                                        framework::Tensor* matrix_out,
-                                        double beta,
-                                        platform::DeviceContext* context) {
+void matmul<platform::GPUPlace, double>(
+    const platform::DeviceContext& context, const framework::Tensor& matrix_a,
+    bool trans_a, const framework::Tensor& matrix_b, bool trans_b, double alpha,
+    framework::Tensor* matrix_out, double beta) {
   auto dim_a = matrix_a.dims();
   auto dim_b = matrix_b.dims();
   auto dim_out = matrix_out->dims();
@@ -118,8 +115,8 @@ void matmul<platform::GPUPlace, double>(const framework::Tensor& matrix_a,
   CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans;
 
   gemm<platform::GPUPlace, double>(
-      transA, transB, M, N, K, alpha, matrix_a.data<double>(),
-      matrix_b.data<double>(), beta, matrix_out->data<double>(), context);
+      context, transA, transB, M, N, K, alpha, matrix_a.data<double>(),
+      matrix_b.data<double>(), beta, matrix_out->data<double>());
 }
 
 }  // namespace math

paddle/operators/math/math_function.h

@@ -66,16 +66,16 @@ namespace math {
 // For more detailed info, please refer to
 // http://www.netlib.org/lapack/explore-html/d4/de2/sgemm_8f.html
 template <typename Place, typename T>
-void gemm(const CBLAS_TRANSPOSE transA, const CBLAS_TRANSPOSE transB,
-          const int M, const int N, const int K, const T alpha, const T* A,
-          const T* B, const T beta, T* C, platform::DeviceContext* context);
+void gemm(const platform::DeviceContext& context, const CBLAS_TRANSPOSE transA,
+          const CBLAS_TRANSPOSE transB, const int M, const int N, const int K,
+          const T alpha, const T* A, const T* B, const T beta, T* C);
 
 // matrix multiply with continuous memory
 template <typename Place, typename T>
-void matmul(const framework::Tensor& matrix_a, bool trans_a,
+void matmul(const platform::DeviceContext& context,
+            const framework::Tensor& matrix_a, bool trans_a,
             const framework::Tensor& matrix_b, bool trans_b, T alpha,
-            framework::Tensor* matrix_out, T beta,
-            platform::DeviceContext* context);
+            framework::Tensor* matrix_out, T beta);
 
 }  // namespace math
 }  // namespace operators

paddle/operators/math/math_function_test.cc

@@ -15,8 +15,7 @@ TEST(math_function, notrans_mul_trans) {
   memcpy(input1_ptr, arr, 6 * sizeof(float));
 
   auto* gpu_place = new paddle::platform::GPUPlace(0);
-  paddle::platform::DeviceContext* context =
-      new paddle::platform::CUDADeviceContext(*gpu_place);
+  paddle::platform::CUDADeviceContext context(*gpu_place);
 
   input1_gpu.CopyFrom<float>(input1, *gpu_place);
   input2_gpu.CopyFrom<float>(input1, *gpu_place);
@@ -24,7 +23,7 @@ TEST(math_function, notrans_mul_trans) {
   out_gpu.mutable_data<float>({2, 2}, *gpu_place);
 
   paddle::operators::math::matmul<paddle::platform::GPUPlace, float>(
-      input1_gpu, false, input2_gpu, true, 1, &out_gpu, 0, context);
+      context, input1_gpu, false, input2_gpu, true, 1, &out_gpu, 0);
 
   out.CopyFrom<float>(out_gpu, *cpu_place);
 
@@ -33,6 +32,7 @@ TEST(math_function, notrans_mul_trans) {
   EXPECT_EQ(out_ptr[1], 14);
   EXPECT_EQ(out_ptr[2], 14);
   EXPECT_EQ(out_ptr[3], 50);
+  delete gpu_place;
 }
 
 TEST(math_function, trans_mul_notrans) {
@@ -48,8 +48,7 @@ TEST(math_function, trans_mul_notrans) {
   memcpy(input1_ptr, arr, 6 * sizeof(float));
 
   auto* gpu_place = new paddle::platform::GPUPlace(0);
-  paddle::platform::DeviceContext* context =
-      new paddle::platform::CUDADeviceContext(*gpu_place);
+  paddle::platform::CUDADeviceContext context(*gpu_place);
 
   input1_gpu.CopyFrom<float>(input1, *gpu_place);
   input2_gpu.CopyFrom<float>(input1, *gpu_place);
@@ -57,7 +56,7 @@ TEST(math_function, trans_mul_notrans) {
   out_gpu.mutable_data<float>({3, 3}, *gpu_place);
 
   paddle::operators::math::matmul<paddle::platform::GPUPlace, float>(
-      input1_gpu, true, input2_gpu, false, 1, &out_gpu, 0, context);
+      context, input1_gpu, true, input2_gpu, false, 1, &out_gpu, 0);
 
   out.CopyFrom<float>(out_gpu, *cpu_place);
 
@@ -71,5 +70,6 @@ TEST(math_function, trans_mul_notrans) {
   EXPECT_EQ(out_ptr[6], 15);
   EXPECT_EQ(out_ptr[7], 22);
   EXPECT_EQ(out_ptr[8], 29);
+  delete gpu_place;
 }
 #endif

paddle/operators/mul_op.h

@@ -46,10 +46,8 @@ class MulKernel : public framework::OpKernel {
             : *y;
 
     z->mutable_data<T>(context.GetPlace());
-    auto* device_context =
-        const_cast<platform::DeviceContext*>(context.device_context_);
-    math::matmul<Place, T>(x_matrix, false, y_matrix, false, 1, z, 0,
-                           device_context);
+    math::matmul<Place, T>(context.device_context(), x_matrix, false, y_matrix,
+                           false, 1, z, 0);
   }
 };
 
@@ -71,16 +69,14 @@ class MulGradKernel : public framework::OpKernel {
 
     Tensor* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
     Tensor* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
-    auto* device_context =
-        const_cast<platform::DeviceContext*>(ctx.device_context_);
     if (dx) {
       dx->mutable_data<T>(ctx.GetPlace());
       Tensor dx_matrix = dx->dims().size() > 2 ? framework::ReshapeToMatrix<T>(
                                                      *dx, x_num_col_dims)
                                                : *dx;
       // dx = dout * y'. dx: M x K, dout : M x N, y : K x N
-      math::matmul<Place, T>(*dout, false, y_matrix, true, 1, &dx_matrix, 0,
-                             device_context);
+      math::matmul<Place, T>(ctx.device_context(), *dout, false, y_matrix, true,
+                             1, &dx_matrix, 0);
     }
     if (dy) {
       dy->mutable_data<T>(ctx.GetPlace());
@@ -88,8 +84,8 @@ class MulGradKernel : public framework::OpKernel {
                                                      *dy, y_num_col_dims)
                                                : *dy;
       // dy = x' * dout. dy K x N, dout : M x N, x : M x K
-      math::matmul<Place, T>(x_matrix, true, *dout, false, 1, &dy_matrix, 0,
-                             device_context);
+      math::matmul<Place, T>(ctx.device_context(), x_matrix, true, *dout, false,
+                             1, &dy_matrix, 0);
     }
   }
 };

paddle/platform/CMakeLists.txt

@@ -24,4 +24,4 @@ cc_library(device_context SRCS device_context.cc DEPS memory buddy_allocator
 nv_test(device_context_test SRCS device_context_test.cc DEPS device_context gpu_info)
 
 nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda)
-nv_test(transform_test SRCS transform_test.cu DEPS paddle_memory place)
+nv_test(transform_test SRCS transform_test.cu DEPS paddle_memory place device_context)

paddle/platform/device_context.cc

@@ -101,19 +101,17 @@ CUDADeviceContext::CUDADeviceContext(GPUPlace place) : place_(place) {
   eigen_stream_.reset(new EigenCudaStreamDevice());
   eigen_stream_->Reinitialize(&stream_, place);
   eigen_device_.reset(new Eigen::GpuDevice(eigen_stream_.get()));
+  PADDLE_ENFORCE(dynload::cublasCreate(&cublas_handle_));
+  PADDLE_ENFORCE(dynload::cublasSetStream(cublas_handle_, stream_));
+  PADDLE_ENFORCE(dynload::cudnnCreate(&cudnn_handle_));
+  PADDLE_ENFORCE(dynload::cudnnSetStream(cudnn_handle_, stream_));
 }
 
 CUDADeviceContext::~CUDADeviceContext() {
   SetDeviceId(place_.device);
   Wait();
-  if (cublas_handle_) {
   PADDLE_ENFORCE(dynload::cublasDestroy(cublas_handle_));
-  }
-
-  if (cudnn_handle_) {
   PADDLE_ENFORCE(dynload::cudnnDestroy(cudnn_handle_));
-  }
-
   eigen_stream_.reset();
   eigen_device_.reset();
   PADDLE_ENFORCE(cudaStreamDestroy(stream_));
@@ -129,25 +127,13 @@ Eigen::GpuDevice* CUDADeviceContext::eigen_device() const {
   return eigen_device_.get();
 }
 
-cublasHandle_t CUDADeviceContext::cublas_handle() {
-  if (!cublas_handle_) {
-    SetDeviceId(place_.device);
-    PADDLE_ENFORCE(dynload::cublasCreate(&cublas_handle_));
-    PADDLE_ENFORCE(dynload::cublasSetStream(cublas_handle_, stream_));
-  }
+cublasHandle_t CUDADeviceContext::cublas_handle() const {
   return cublas_handle_;
 }
 
-cudnnHandle_t CUDADeviceContext::cudnn_handle() {
-  if (!cudnn_handle_) {
-    SetDeviceId(place_.device);
-    PADDLE_ENFORCE(dynload::cudnnCreate(&cudnn_handle_));
-    PADDLE_ENFORCE(dynload::cudnnSetStream(cudnn_handle_, stream_));
-  }
-  return cudnn_handle_;
-}
+cudnnHandle_t CUDADeviceContext::cudnn_handle() const { return cudnn_handle_; }
 
-cudaStream_t CUDADeviceContext::stream() { return stream_; }
+cudaStream_t CUDADeviceContext::stream() const { return stream_; }
 
 #endif  // PADDLE_ONLY_CPU
 

paddle/platform/device_context.h

@@ -67,16 +67,14 @@ class CUDADeviceContext : public DeviceContext {
   /*! \brief  Return eigen device in the device context. */
   Eigen::GpuDevice* eigen_device() const;
 
-  // clang-format off
   /*! \brief  Return cublas handle in the device context. */
-  cublasHandle_t    cublas_handle();
+  cublasHandle_t cublas_handle() const;
 
   /*! \brief  Return cudnn  handle in the device context. */
-  cudnnHandle_t     cudnn_handle();
+  cudnnHandle_t cudnn_handle() const;
 
   /*! \brief  Return cuda stream in the device context. */
-  cudaStream_t      stream();
-  // clang-format on
+  cudaStream_t stream() const;
 
  private:
   GPUPlace place_;
@@ -84,11 +82,9 @@ class CUDADeviceContext : public DeviceContext {
   std::unique_ptr<Eigen::GpuDevice> eigen_device_;
   std::unique_ptr<EigenCudaStreamDevice> eigen_stream_;
 
-  // clang-format off
-  cudaStream_t       stream_{nullptr};
-  cudnnHandle_t      cudnn_handle_{nullptr};
-  cublasHandle_t     cublas_handle_{nullptr};
-  // clang-format on
+  cudaStream_t stream_;
+  cudnnHandle_t cudnn_handle_;
+  cublasHandle_t cublas_handle_;
 };
 
 #endif

paddle/platform/transform.h

@@ -14,6 +14,7 @@
 
 #pragma once
 
+#include "paddle/platform/device_context.h"
 #include "paddle/platform/enforce.h"
 #include "paddle/platform/hostdevice.h"
 #include "paddle/platform/place.h"
@@ -21,6 +22,7 @@
 #include <algorithm>
 #include <type_traits>
 #ifdef __NVCC__
+#include <thrust/execution_policy.h>
 #include <thrust/transform.h>
 #include "paddle/platform/details/device_ptr_cast.h"
 #endif
@@ -28,34 +30,39 @@
 namespace paddle {
 namespace platform {
 // Transform on host or device. It provides the same API in std library.
-template <typename Place, typename InputIter, typename OutputIter,
-          typename UnaryOperation>
-void Transform(Place place, InputIter first, InputIter last, OutputIter result,
-               UnaryOperation op) {
+template <typename InputIter, typename OutputIter, typename UnaryOperation>
+void Transform(const DeviceContext& context, InputIter first, InputIter last,
+               OutputIter result, UnaryOperation op) {
+  auto place = context.GetPlace();
   if (is_cpu_place(place)) {
     std::transform(first, last, result, op);
   } else {
 #ifdef __NVCC__
+    auto& ctx = reinterpret_cast<const CUDADeviceContext&>(context);
     using namespace details;
-    thrust::transform(DevPtrCast(first), DevPtrCast(last), DevPtrCast(result),
-                      op);
+    thrust::transform(thrust::cuda::par.on(ctx.stream()), DevPtrCast(first),
+                      DevPtrCast(last), DevPtrCast(result), op);
 #else
     PADDLE_THROW("Do not invoke `Transform<GPUPlace>` in .cc file");
 #endif
   }
 }
 
-template <typename Place, typename InputIter1, typename InputIter2,
-          typename OutputIter, typename BinaryOperation>
-void Transform(Place place, InputIter1 first1, InputIter1 last1,
-               InputIter2 first2, OutputIter result, BinaryOperation op) {
+template <typename InputIter1, typename InputIter2, typename OutputIter,
+          typename BinaryOperation>
+void Transform(const DeviceContext& context, InputIter1 first1,
+               InputIter1 last1, InputIter2 first2, OutputIter result,
+               BinaryOperation op) {
+  auto place = context.GetPlace();
   if (is_cpu_place(place)) {
     std::transform(first1, last1, first2, result, op);
   } else {
 #ifdef __NVCC__
+    auto& ctx = reinterpret_cast<const CUDADeviceContext&>(context);
     using namespace details;
-    thrust::transform(DevPtrCast(first1), DevPtrCast(last1), DevPtrCast(first2),
-                      DevPtrCast(result), op);
+    thrust::transform(thrust::cuda::par.on(ctx.stream()), DevPtrCast(first1),
+                      DevPtrCast(last1), DevPtrCast(first2), DevPtrCast(result),
+                      op);
 #else
     PADDLE_THROW("Do not invoke `Transform<GPUPlace>` in .cc file");
 #endif

paddle/platform/transform_test.cu

@@ -36,8 +36,9 @@ class Multiply {
 
 TEST(Transform, CPUUnary) {
   using namespace paddle::platform;
+  CPUDeviceContext ctx;
   float buf[4] = {0.1, 0.2, 0.3, 0.4};
-  Transform(CPUPlace(), buf, buf + 4, buf, Scale<float>(10));
+  Transform(ctx, buf, buf + 4, buf, Scale<float>(10));
   for (int i = 0; i < 4; ++i) {
     ASSERT_NEAR(buf[i], static_cast<float>(i + 1), 1e-5);
   }
@@ -47,10 +48,12 @@ TEST(Transform, GPUUnary) {
   using namespace paddle::platform;
   using namespace paddle::memory;
   GPUPlace gpu0(0);
+  CUDADeviceContext ctx(gpu0);
   float cpu_buf[4] = {0.1, 0.2, 0.3, 0.4};
   float* gpu_buf = static_cast<float*>(Alloc(gpu0, sizeof(float) * 4));
   Copy(gpu0, gpu_buf, CPUPlace(), cpu_buf, sizeof(cpu_buf));
-  Transform(gpu0, gpu_buf, gpu_buf + 4, gpu_buf, Scale<float>(10));
+  Transform(ctx, gpu_buf, gpu_buf + 4, gpu_buf, Scale<float>(10));
+  ctx.Wait();
   Copy(CPUPlace(), cpu_buf, gpu0, gpu_buf, sizeof(cpu_buf));
   Free(gpu0, gpu_buf);
   for (int i = 0; i < 4; ++i) {
@@ -62,7 +65,7 @@ TEST(Transform, CPUBinary) {
   using namespace paddle::platform;
   using namespace paddle::memory;
   int buf[4] = {1, 2, 3, 4};
-  Transform(CPUPlace(), buf, buf + 4, buf, buf, Multiply<int>());
+  Transform(CPUDeviceContext(), buf, buf + 4, buf, buf, Multiply<int>());
   for (int i = 0; i < 4; ++i) {
     ASSERT_EQ((i + 1) * (i + 1), buf[i]);
   }
@@ -73,9 +76,11 @@ TEST(Transform, GPUBinary) {
   using namespace paddle::memory;
   int buf[4] = {1, 2, 3, 4};
   GPUPlace gpu0(0);
+  CUDADeviceContext ctx(gpu0);
   int* gpu_buf = static_cast<int*>(Alloc(gpu0, sizeof(buf)));
   Copy(gpu0, gpu_buf, CPUPlace(), buf, sizeof(buf));
-  Transform(gpu0, gpu_buf, gpu_buf + 4, gpu_buf, gpu_buf, Multiply<int>());
+  Transform(ctx, gpu_buf, gpu_buf + 4, gpu_buf, gpu_buf, Multiply<int>());
+  ctx.Wait();
   Copy(CPUPlace(), buf, gpu0, gpu_buf, sizeof(buf));
   Free(gpu0, gpu_buf);
   for (int i = 0; i < 4; ++i) {

python/paddle/trainer/config_parser.py

@@ -2286,8 +2286,15 @@ class NormLayer(LayerBase):
 
 @config_layer('pool')
 class PoolLayer(LayerBase):
+    layer_type = 'pool'
+
     def __init__(self, name, inputs, ceil_mode=True, **xargs):
-        super(PoolLayer, self).__init__(name, 'pool', 0, inputs=inputs, **xargs)
+        use_mkldnn = int(g_command_config_args.get("use_mkldnn", 0))
+        if self.layer_type == "mkldnn_pool":
+            config_assert(use_mkldnn, "mkldnn_pool only support MKLDNN")
+        self.layer_type = 'mkldnn_pool' if use_mkldnn else 'pool'
+        super(PoolLayer, self).__init__(
+            name, self.layer_type, 0, inputs=inputs, **xargs)
         for input_index in xrange(len(self.inputs)):
             input_layer = self.get_input_layer(input_index)
             pool_conf = self.config.inputs[input_index].pool_conf
@@ -2297,6 +2304,11 @@ class PoolLayer(LayerBase):
                                pool_conf.channels)
 
 
+@config_layer('mkldnn_pool')
+class MKLDNNPoolLayer(PoolLayer):
+    layer_type = 'mkldnn_pool'
+
+
 @config_layer('pool3d')
 class Pool3DLayer(LayerBase):
     def __init__(self, name, inputs, ceil_mode=True, **xargs):

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

@@ -192,6 +192,9 @@ class OpTest(unittest.TestCase):
         self.op.run(self.scope, ctx)
 
         for out_name, out_dup in Operator.get_op_outputs(self.op.type()):
+            if out_name not in self.outputs:
+                continue
+
             if out_dup:
                 sub_out = self.outputs[out_name]
                 if not isinstance(sub_out, list):
@@ -206,9 +209,7 @@ class OpTest(unittest.TestCase):
                             actual, expect, atol=1e-05),
                         "output name: " + out_name + " has diff")
             else:
-                var = self.scope.find_var(out_name)
-                if var is not None:
-                    actual = np.array(var.get_tensor())
+                actual = np.array(self.scope.find_var(out_name).get_tensor())
                 expect = self.outputs[out_name]
                 self.assertTrue(
                     np.allclose(