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

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/paddle/image/alexnet.py

@@ -6,8 +6,18 @@ height = 227
 width = 227
 num_class = 1000
 batch_size = get_config_arg('batch_size', int, 128)
+gp = get_config_arg('layer_num', int, 1)
+is_infer = get_config_arg("is_infer", bool, False)
+num_samples = get_config_arg('num_samples', int, 2560)
 
-args = {'height': height, 'width': width, 'color': True, 'num_class': num_class}
+args = {
+    'height': height,
+    'width': width,
+    'color': True,
+    'num_class': num_class,
+    'is_infer': is_infer,
+    'num_samples': num_samples
+}
 define_py_data_sources2(
     "train.list", None, module="provider", obj="process", args=args)
 
@@ -31,7 +41,7 @@ net = img_pool_layer(input=net, pool_size=3, stride=2)
 
 # conv2
 net = img_conv_layer(
-    input=net, filter_size=5, num_filters=256, stride=1, padding=2, groups=1)
+    input=net, filter_size=5, num_filters=256, stride=1, padding=2, groups=gp)
 net = img_cmrnorm_layer(input=net, size=5, scale=0.0001, power=0.75)
 net = img_pool_layer(input=net, pool_size=3, stride=2)
 
@@ -40,11 +50,11 @@ net = img_conv_layer(
     input=net, filter_size=3, num_filters=384, stride=1, padding=1)
 # conv4
 net = img_conv_layer(
-    input=net, filter_size=3, num_filters=384, stride=1, padding=1, groups=1)
+    input=net, filter_size=3, num_filters=384, stride=1, padding=1, groups=gp)
 
 # conv5
 net = img_conv_layer(
-    input=net, filter_size=3, num_filters=256, stride=1, padding=1, groups=1)
+    input=net, filter_size=3, num_filters=256, stride=1, padding=1, groups=gp)
 net = img_pool_layer(input=net, pool_size=3, stride=2)
 
 net = fc_layer(
@@ -59,6 +69,9 @@ net = fc_layer(
     layer_attr=ExtraAttr(drop_rate=0.5))
 net = fc_layer(input=net, size=1000, act=SoftmaxActivation())
 
-lab = data_layer('label', num_class)
-loss = cross_entropy(input=net, label=lab)
-outputs(loss)
+if is_infer:
+    outputs(net)
+else:
+    lab = data_layer('label', num_class)
+    loss = cross_entropy(input=net, label=lab)
+    outputs(loss)

benchmark/paddle/image/googlenet.py

@@ -7,13 +7,15 @@ num_class = 1000
 batch_size = get_config_arg('batch_size', int, 128)
 use_gpu = get_config_arg('use_gpu', bool, True)
 is_infer = get_config_arg("is_infer", bool, False)
+num_samples = get_config_arg('num_samples', int, 2560)
 
 args = {
     'height': height,
     'width': width,
     'color': True,
     'num_class': num_class,
-    'is_infer': is_infer
+    'is_infer': is_infer,
+    'num_samples': num_samples
 }
 define_py_data_sources2(
     "train.list" if not is_infer else None,

benchmark/paddle/image/provider.py

@@ -14,6 +14,7 @@ def initHook(settings, height, width, color, num_class, **kwargs):
     else:
         settings.data_size = settings.height * settings.width
     settings.is_infer = kwargs.get('is_infer', False)
+    settings.num_samples = kwargs.get('num_samples', 2560)
     if settings.is_infer:
         settings.slots = [dense_vector(settings.data_size)]
     else:
@@ -23,7 +24,7 @@ def initHook(settings, height, width, color, num_class, **kwargs):
 @provider(
     init_hook=initHook, min_pool_size=-1, cache=CacheType.CACHE_PASS_IN_MEM)
 def process(settings, file_list):
-    for i in xrange(2560 if settings.is_infer else 1024):
+    for i in xrange(settings.num_samples):
         img = np.random.rand(1, settings.data_size).reshape(-1, 1).flatten()
         if settings.is_infer:
             yield img.astype('float32')

benchmark/paddle/image/resnet.py

@@ -7,13 +7,15 @@ num_class = 1000
 batch_size = get_config_arg('batch_size', int, 64)
 layer_num = get_config_arg("layer_num", int, 50)
 is_infer = get_config_arg("is_infer", bool, False)
+num_samples = get_config_arg('num_samples', int, 2560)
 
 args = {
     'height': height,
     'width': width,
     'color': True,
     'num_class': num_class,
-    'is_infer': is_infer
+    'is_infer': is_infer,
+    'num_samples': num_samples
 }
 define_py_data_sources2(
     "train.list" if not is_infer else None,

benchmark/paddle/image/run_mkl_infer.sh

@@ -37,7 +37,7 @@ function infer() {
       --trainer_count=1 \
       --num_passes=1 \
       --save_dir="models/${topology}-${layer_num}" \
-      --config_args="batch_size=128,layer_num=${layer_num}" \
+      --config_args="batch_size=128,layer_num=${layer_num},num_samples=256" \
       > /dev/null 2>&1
     echo "Done"
   fi
@@ -79,8 +79,9 @@ fi
 # inference benchmark
 for use_mkldnn in True False; do
   for batchsize in 1 2 4 8 16; do
-    infer googlenet v1 $batchsize $use_mkldnn
-    infer resnet 50 $batchsize $use_mkldnn
     infer vgg 19 $batchsize $use_mkldnn
+    infer resnet 50 $batchsize $use_mkldnn
+    infer googlenet v1 $batchsize $use_mkldnn
+    infer alexnet 2 $batchsize $use_mkldnn
   done
 done

benchmark/paddle/image/run_mkl_train.sh

@@ -47,5 +47,6 @@ for use_mkldnn in True False; do
     train vgg 19 $batchsize $use_mkldnn
     train resnet 50 $batchsize $use_mkldnn
     train googlenet v1 $batchsize $use_mkldnn
+    train alexnet 2 $batchsize $use_mkldnn
   done
 done

benchmark/paddle/image/run_openblas_infer.sh

@@ -23,24 +23,25 @@ function infer() {
     echo "./run_mkl_infer.sh to save the model first"
     exit 0
   fi
-  log_period=$((256 / bs))
+  log_period=$((32 / bs))
   paddle train --job=test \
     --config="${topology}.py" \
+    --use_mkldnn=False \
     --use_gpu=False \
     --trainer_count=$thread \
     --log_period=$log_period \
-    --config_args="batch_size=${bs},layer_num=${layer_num},is_infer=True" \
+    --config_args="batch_size=${bs},layer_num=${layer_num},is_infer=True,num_samples=256" \
     --init_model_path=$models_in \
     2>&1 | tee ${log}
 
-  # calculate the last 5 logs period time of 1280 samples,
+  # calculate the last 5 logs period time of 160(=32*5) samples,
   # the time before are burning time.
   start=`tail ${log} -n 7 | head -n 1 | awk -F ' ' '{print $2}' | xargs`
   end=`tail ${log} -n 2 | head -n 1 | awk -F ' ' '{print $2}' | xargs`
   start_sec=`clock_to_seconds $start`
   end_sec=`clock_to_seconds $end`
-  fps=`awk 'BEGIN{printf "%.2f",(1280 / ('$end_sec' - '$start_sec'))}'`
-  echo "Last 1280 samples start: ${start}(${start_sec} sec), end: ${end}(${end_sec} sec;" >> ${log}
+  fps=`awk 'BEGIN{printf "%.2f",(160 / ('$end_sec' - '$start_sec'))}'`
+  echo "Last 160 samples start: ${start}(${start_sec} sec), end: ${end}(${end_sec} sec;" >> ${log}
   echo "FPS: $fps images/sec" 2>&1 | tee -a ${log}
 }
 
@@ -56,7 +57,8 @@ fi
 
 # inference benchmark
 for batchsize in 1 2 4 8 16; do
-  infer googlenet v1 $batchsize
-  infer resnet 50 $batchsize
   infer vgg 19 $batchsize
+  infer resnet 50 $batchsize 
+  infer googlenet v1 $batchsize
+  infer alexnet 2 $batchsize
 done

benchmark/paddle/image/run_openblas_train.sh

@@ -12,10 +12,11 @@ function train() {
   config="${topology}.py"
   paddle train --job=time \
     --config=$config \
+    --use_mkldnn=False \
     --use_gpu=False \
     --trainer_count=$thread \
-    --log_period=10 \
-    --test_period=100 \
+    --log_period=3 \
+    --test_period=30 \
     --config_args=$args \
     2>&1 | tee ${log} 
 
@@ -36,4 +37,5 @@ for batchsize in 64 128 256; do
   train vgg 19 $batchsize
   train resnet 50 $batchsize
   train googlenet v1 $batchsize
+  train alexnet 2 $batchsize
 done

benchmark/paddle/image/vgg.py

@@ -7,13 +7,15 @@ num_class = 1000
 batch_size = get_config_arg('batch_size', int, 64)
 layer_num = get_config_arg('layer_num', int, 19)
 is_infer = get_config_arg("is_infer", bool, False)
+num_samples = get_config_arg('num_samples', int, 2560)
 
 args = {
     'height': height,
     'width': width,
     'color': True,
     'num_class': num_class,
-    'is_infer': is_infer
+    'is_infer': is_infer,
+    'num_samples': num_samples
 }
 define_py_data_sources2(
     "train.list" if not is_infer else None,

doc/api/v2/fluid/layers.rst

@@ -170,6 +170,18 @@ 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
@@ -318,3 +330,9 @@ reduce_sum
 ..  autofunction:: paddle.v2.fluid.layers.reduce_sum
     :noindex:
 
+
+reduce_mean
+---------
+..  autofunction:: paddle.v2.fluid.layers.reduce_mean
+    :noindex:
+

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/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/getstarted/build_and_install/build_from_source_cn.rst

@@ -70,13 +70,13 @@ PaddlePaddle编译需要使用到下面的依赖（包含但不限于），其
    :header: "依赖", "版本", "说明"
    :widths: 10, 15, 30
 
-   "CMake", ">=3.5", ""
+   "CMake", ">=3.2", ""
    "GCC", "4.8.2", "推荐使用CentOS的devtools2"
-   "Python", "2.7.x", "依赖libpython2.7.so"
-   "pip", ">=9.0", ""
-   "numpy", "", ""
+   "Python", "2.7.x", "依赖libpython2.7.so"
+   "pip", ">=9.0", ""
+   "numpy", "", ""
    "SWIG", ">=2.0", ""
-   "Go", ">=1.8", "可选"
+   "Go", ">=1.8", "可选"
 
 
 .. _build_options:

doc/getstarted/build_and_install/build_from_source_en.rst

@@ -76,13 +76,13 @@ will be downloaded automatically.
    :header: "Dependency", "Version", "Description"
    :widths: 10, 15, 30
 
-   "CMake", ">=3.5", ""
+   "CMake", ">=3.2", ""
    "GCC", "4.8.2", "Recommend devtools2 for CentOS"
-   "Python", "2.7.x", "Need libpython2.7.so"
-   "pip", ">=9.0", ""
-   "numpy", "", ""
+   "Python", "2.7.x", "Need libpython2.7.so"
+   "pip", ">=9.0", ""
+   "numpy", "", ""
    "SWIG", ">=2.0", ""
-   "Go", ">=1.8", "Optional"
+   "Go", ">=1.8", "Optional"
 
 
 .. _build_options:

paddle/framework/CMakeLists.txt

@@ -30,7 +30,7 @@ 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_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)
@@ -59,5 +59,8 @@ 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_test(threadpool_test SRCS threadpool_test.cc)
+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)

paddle/framework/block_desc.cc

@@ -90,6 +90,21 @@ OpDesc *BlockDesc::PrependOp() {
   return ops_.front().get();
 }
 
+void BlockDesc::RemoveOp(size_t s, size_t e) {
+  if (ops_.begin() + s == ops_.end() || ops_.begin() + e == ops_.end()) {
+    return;
+  }
+  need_update_ = true;
+  for (auto it = ops_.begin() + s; it != ops_.begin() + e; it++) {
+    auto names = (*it)->InputArgumentNames();
+    for (auto n : names) {
+      // TODO(typhoonzero): delete vars if no other op use it.
+      VLOG(3) << "deleting var " << n;
+    }
+  }
+  ops_.erase(ops_.begin() + s, ops_.begin() + e);
+}
+
 std::vector<OpDesc *> BlockDesc::AllOps() const {
   std::vector<OpDesc *> res;
   for (const auto &op : ops_) {

paddle/framework/block_desc.h

@@ -79,6 +79,8 @@ class BlockDesc {
 
   OpDesc *PrependOp();
 
+  void RemoveOp(size_t s, size_t e);
+
   std::vector<OpDesc *> AllOps() const;
 
   size_t OpSize() const { return ops_.size(); }

paddle/framework/data_layout.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 <iostream>
+#include "paddle/platform/enforce.h"
+
+namespace paddle {
+namespace framework {
+
+enum DataLayout {
+  kNHWC = 0,
+  kNCHW = 1,
+  kAnyLayout = 2,
+};
+
+inline DataLayout StringToDataLayout(const std::string& str) {
+  if (str == "NHWC" || str == "nhwc") {
+    return DataLayout::kNHWC;
+  } else if (str == "NCHW" || str == "nchw") {
+    return DataLayout::kNCHW;
+  } else {
+    PADDLE_THROW("Unknown storage order string: %s", str);
+  }
+}
+
+inline std::string DataLayoutToString(const DataLayout& data_layout) {
+  switch (data_layout) {
+    case kNHWC:
+      return "NHWC";
+    case kNCHW:
+      return "NCHW";
+    case 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/executor.cc

@@ -33,13 +33,7 @@ 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) {
@@ -65,15 +59,15 @@ static void CreateTensor(Variable* var, proto::VarDesc::VarType var_type) {
 }
 
 void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
-                   bool create_local_scope) {
+                   bool create_local_scope, bool create_vars) {
   // TODO(tonyyang-svail):
   //    - only runs on the first device (i.e. no interdevice communication)
   //    - 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) {
     if (create_local_scope) {
       local_scope = &scope->NewScope();
       for (auto& var : block.AllVars()) {
@@ -100,12 +94,13 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
         VLOG(3) << "Create variable " << var->Name() << ", which pointer is "
                 << ptr;
       }
-  }
+    }  // if (create_local_scope)
+  }    // if (create_vars)
 
   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 (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"
@@ -26,86 +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;
-  }
-
-  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
@@ -114,10 +38,11 @@ class Executor {
    *  ProgramDesc
    *  Scope
    */
-  void Run(const ProgramDesc&, Scope*, int, bool create_local_scope = true);
+  void Run(const ProgramDesc&, Scope*, int, bool create_local_scope = true,
+           bool create_vars = true);
 
  private:
-  std::vector<const platform::DeviceContext*> device_contexts_;
+  const platform::Place place_;
 };
 
 }  // namespace framework

paddle/framework/grad_op_desc_maker.h

@@ -22,6 +22,14 @@
 namespace paddle {
 namespace framework {
 
+/*
+  This functor class is responsible for creating the gradient ops for the given
+  operator fwd_op. After it is called (through operator()), the pairs of
+  (gradient variable, corresponding input variable of fwd_op) will be added to
+  grad_to_var. If an input variable of fwd_op is contained in no_grad_set, its
+  gradient varialbe will be ignored or kEmptyVarName depending on the template
+  argument DropEmptyIG in the derived classes.
+ */
 class GradOpDescMakerBase {
  public:
   explicit GradOpDescMakerBase(
@@ -56,6 +64,16 @@ class GradOpDescMakerBase {
     if (!drop_empty_grad) {
       return ret_val;
     }
+    PADDLE_ENFORCE_LE(var_names.size(), 1UL,
+                      "BUG from operator developer:"
+                      " for input argument with a list of variables, "
+                      " drop_empty_grad is not allowed because it makes"
+                      " the correspondence bewteen a variable and its gradient"
+                      " ambiguous. Use REGISTER_OP_EX to register the op"
+                      " or call InputGrad(?,false) in GradOpDescMaker."
+                      " Op type %s",
+                      fwd_op_.Type());
+
     std::vector<std::string> dropped_ret_val;
     dropped_ret_val.reserve(ret_val.size());
     std::copy_if(ret_val.begin(), ret_val.end(),

paddle/framework/init.cc

@@ -14,8 +14,8 @@
 #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::Create(places);
   return true;
 }
 

paddle/framework/init_test.cc

@@ -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

+/* 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
+
+namespace paddle {
+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 };
+
+inline std::string LibraryTypeToString(const LibraryType& library_type) {
+  switch (library_type) {
+    case kPlain:
+      return "PLAIN";
+    case kMKLDNN:
+      return "MKLDNN";
+    case kCUDNN:
+      return "CUDNN";
+    default:
+      PADDLE_THROW("unknown LibraryType %d", library_type);
+  }
+}
+
+inline std::ostream& operator<<(std::ostream& out, LibraryType l) {
+  out << LibraryTypeToString(l);
+  return out;
+}
+
+}  // namespace
+}  // framework

paddle/framework/lod_tensor.cc

@@ -224,7 +224,7 @@ void SerializeToStream(std::ostream &os, const LoDTensor &tensor,
       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()),
+                     boost::get<platform::CUDAPlace>(tensor.place()),
                      reinterpret_cast<const void *>(data), size_to_write,
                      gpu_dev_ctx.stream());
         gpu_dev_ctx.Wait();

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.h

@@ -127,7 +127,9 @@ class OpDesc {
   }
 
   proto::OpDesc desc_;
+  // input arg name => output variable names
   VariableNameMap inputs_;
+  // output arg name => output variable names
   VariableNameMap outputs_;
   AttributeMap attrs_;
 

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_;
+  }
+};
+
+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;
+}
+
+}  // 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);
+
+  std::ostringstream stream;
+  stream << op_kernel_type;
+  ASSERT_EQ(
+      stream.str(),
+      "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));
+}
\ No newline at end of file

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,
@@ -126,6 +115,14 @@ class OpKernelRegistrar : public Registrar {
                              __test_global_namespace_##uniq_name##__>::value, \
                 msg)
 
+/*
+  The variadic arguments should be class types derived from one of the
+  following classes:
+    OpProtoAndCheckerMaker
+    GradOpDescMakerBase
+    VarTypeInference
+    InferShapeBase
+*/
 #define REGISTER_OPERATOR(op_type, op_class, ...)                      \
   STATIC_ASSERT_GLOBAL_NAMESPACE(                                      \
       __reg_op__##op_type,                                             \
@@ -144,15 +141,24 @@ class OpKernelRegistrar : public Registrar {
   }
 
 /**
- * Macro to register Operator.
+ * Macro to register Operator. When the input is duplicable, you should
+ * use REGISTER_OP_EX with deop_empty_grad=false instead.
  */
 #define REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type, \
                     grad_op_class)                                   \
+  REGISTER_OP_EX(op_type, op_class, op_maker_class, grad_op_type,    \
+                 grad_op_class, true)
+
+// When an argument is duplicable, we need to use this version.
+// Perhaps we can omit DropEmptyIG template parameter and
+// only have one version of REGISTER_OP.
+#define REGISTER_OP_EX(op_type, op_class, op_maker_class, grad_op_type,       \
+                       grad_op_class, drop_empty_grad)                        \
   REGISTER_OPERATOR(grad_op_type, grad_op_class);                             \
   class _GradOpDescMaker_##grad_op_type##_                                    \
-      : public ::paddle::framework::DefaultGradOpDescMaker<true> {         \
+      : public ::paddle::framework::DefaultGradOpDescMaker<drop_empty_grad> { \
     using ::paddle::framework::DefaultGradOpDescMaker<                        \
-        true>::DefaultGradOpDescMaker;                                     \
+        drop_empty_grad>::DefaultGradOpDescMaker;                             \
                                                                               \
    protected:                                                                 \
     virtual std::string GradOpType() const { return #grad_op_type; }          \
@@ -182,7 +188,7 @@ class OpKernelRegistrar : public Registrar {
   }
 
 #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

@@ -8,8 +8,7 @@ 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 +27,7 @@ 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 +74,8 @@ 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 +115,8 @@ 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 +165,9 @@ 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);
 }

paddle/framework/operator.cc

@@ -12,10 +12,12 @@ 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/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 +242,6 @@ 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);
@@ -388,11 +384,11 @@ class RuntimeInferShapeContext : public InferShapeContext {
 };
 
 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::Get();
+  auto dev_ctx = pool.Borrow(place);
 
   // check if op[type] has kernel registered.
   auto& all_op_kernels = AllOpKernels();
@@ -404,19 +400,30 @@ 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);
   }
 
   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 +87,7 @@ 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;
 
   virtual bool IsNetOp() const { return false; }
 
@@ -159,8 +162,7 @@ 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 +347,6 @@ 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 +357,7 @@ 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 +378,9 @@ 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 +388,6 @@ 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/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;

paddle/framework/tensor_test.cc

@@ -80,20 +80,20 @@ TEST(Tensor, MutableData) {
     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>(make_ddim({1, 2, 3}), 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>(make_ddim({3, 4}), 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>(make_ddim({2, 2, 3}), 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>(make_ddim({2, 2}), CUDAPlace());
     EXPECT_EQ(p1, p2);
   }
 #endif
@@ -130,7 +130,7 @@ TEST(Tensor, ShareDataWith) {
   {
     Tensor src_tensor;
     Tensor dst_tensor;
-    src_tensor.mutable_data<int>(make_ddim({2, 3, 4}), GPUPlace());
+    src_tensor.mutable_data<int>(make_ddim({2, 3, 4}), CUDAPlace());
     dst_tensor.ShareDataWith(src_tensor);
     ASSERT_EQ(src_tensor.data<int>(), dst_tensor.data<int>());
   }
@@ -166,7 +166,7 @@ TEST(Tensor, Slice) {
 #ifdef PADDLE_WITH_CUDA
   {
     Tensor src_tensor;
-    src_tensor.mutable_data<double>(make_ddim({6, 9}), GPUPlace());
+    src_tensor.mutable_data<double>(make_ddim({6, 9}), CUDAPlace());
     Tensor slice_tensor = src_tensor.Slice(2, 6);
     DDim slice_dims = slice_tensor.dims();
     ASSERT_EQ(arity(slice_dims), 2);
@@ -176,11 +176,11 @@ TEST(Tensor, Slice) {
     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()));
+        src_tensor.mutable_data<double>(src_tensor.dims(), 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()));
+        slice_tensor.mutable_data<double>(slice_tensor.dims(), 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);

paddle/framework/tensor_util.h

@@ -47,11 +47,11 @@ 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 +59,21 @@ 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 +82,28 @@ 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());
+
+  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 +130,28 @@ 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 +178,30 @@ 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);
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/tensor_util_test.cc

@@ -17,6 +17,7 @@
 
 namespace paddle {
 namespace framework {
+
 TEST(CopyFrom, Tensor) {
   Tensor src_tensor;
   Tensor dst_tensor;
@@ -29,7 +30,7 @@ TEST(CopyFrom, Tensor) {
   memcpy(src_ptr, arr, 9 * sizeof(int));
 
   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);
@@ -58,7 +59,7 @@ 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);
 
@@ -104,8 +105,7 @@ 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 +117,7 @@ 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 +143,7 @@ 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 +198,8 @@ 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 +209,7 @@ 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);
 

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 <cstdio>
+#include <functional>
+#include <iostream>
+#include <mutex>
+#include <queue>
+#include <thread>
+
+#include "paddle/platform/call_once.h"
+#include "paddle/platform/enforce.h"
+
+namespace paddle {
+namespace framework {
+
+typedef std::function<void()> Task;
+
+class ThreadPool {
+ public:
+  /**
+   * @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.
+   */
+  void Run(const Task& fn) {
+    std::unique_lock<std::mutex> lock(mutex_);
+    tasks_.push(fn);
+    lock.unlock();
+    scheduled_.notify_one();
+  }
+
+  /**
+   * @brief   Wait until all the tasks are completed.
+   */
+  void Wait() {
+    std::unique_lock<std::mutex> lock(mutex_);
+    completed_.wait(lock, [=] { return Done() == true; });
+  }
+
+ private:
+  ThreadPool& operator=(const ThreadPool&) = delete;
+  ThreadPool(const ThreadPool&) = delete;
+
+  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 = 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_;
+};
+
+std::unique_ptr<ThreadPool> ThreadPool::threadpool(nullptr);
+std::once_flag ThreadPool::init_flag;
+}  // 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 "threadpool.h"
+#include <gtest/gtest.h>
+#include <atomic>
+#include <chrono>
+#include <map>
+#include <thread>
+
+namespace framework = paddle::framework;
+
+void do_sum(framework::ThreadPool* pool, std::atomic<int>& sum, int cnt) {
+  for (int i = 0; i < cnt; ++i) {
+    pool->Run([&sum]() { sum.fetch_add(1); });
+  }
+}
+
+TEST(ThreadPool, ConcurrentInit) {
+  framework::ThreadPool* pool;
+  int concurrent_cnt = 50;
+  std::vector<std::thread> threads;
+  for (int i = 0; i < concurrent_cnt; ++i) {
+    std::thread t([&pool]() { pool = framework::ThreadPool::GetInstance(); });
+    threads.push_back(std::move(t));
+  }
+  for (auto& t : threads) {
+    t.join();
+  }
+}
+
+TEST(ThreadPool, ConcurrentStart) {
+  framework::ThreadPool* pool = framework::ThreadPool::GetInstance();
+  std::atomic<int> sum(0);
+  std::vector<std::thread> threads;
+  int concurrent_cnt = 50;
+  // sum = (n * (n + 1)) / 2
+  for (int i = 1; i <= concurrent_cnt; ++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, ((concurrent_cnt + 1) * concurrent_cnt) / 2);
+}

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/memcpy.cc

@@ -28,31 +28,25 @@ void Copy<platform::CPUPlace, platform::CPUPlace>(platform::CPUPlace, void* dst,
 
 #ifdef PADDLE_WITH_CUDA
 template <>
-void Copy<platform::CPUPlace, platform::GPUPlace>(platform::CPUPlace dst_place,
-                                                  void* dst,
-                                                  platform::GPUPlace src_place,
-                                                  const void* src, size_t num,
-                                                  cudaStream_t stream) {
+void Copy<platform::CPUPlace, platform::CUDAPlace>(
+    platform::CPUPlace dst_place, void* dst, platform::CUDAPlace src_place,
+    const void* src, size_t num, cudaStream_t stream) {
   platform::SetDeviceId(src_place.device);
   platform::GpuMemcpyAsync(dst, src, num, cudaMemcpyDeviceToHost, stream);
 }
 
 template <>
-void Copy<platform::GPUPlace, platform::CPUPlace>(platform::GPUPlace dst_place,
-                                                  void* dst,
-                                                  platform::CPUPlace src_place,
-                                                  const void* src, size_t num,
-                                                  cudaStream_t stream) {
+void Copy<platform::CUDAPlace, platform::CPUPlace>(
+    platform::CUDAPlace dst_place, void* dst, platform::CPUPlace src_place,
+    const void* src, size_t num, cudaStream_t stream) {
   platform::SetDeviceId(dst_place.device);
   platform::GpuMemcpyAsync(dst, src, num, cudaMemcpyHostToDevice, stream);
 }
 
 template <>
-void Copy<platform::GPUPlace, platform::GPUPlace>(platform::GPUPlace dst_place,
-                                                  void* dst,
-                                                  platform::GPUPlace src_place,
-                                                  const void* src, size_t num,
-                                                  cudaStream_t stream) {
+void Copy<platform::CUDAPlace, platform::CUDAPlace>(
+    platform::CUDAPlace dst_place, void* dst, platform::CUDAPlace src_place,
+    const void* src, size_t num, cudaStream_t stream) {
   if (dst_place == src_place) {
     platform::SetDeviceId(src_place.device);
     platform::GpuMemcpyAsync(dst, src, num, cudaMemcpyDeviceToDevice, stream);
@@ -62,33 +56,6 @@ void Copy<platform::GPUPlace, platform::GPUPlace>(platform::GPUPlace dst_place,
   }
 }
 
-template <>
-void Copy<platform::CPUPlace, platform::GPUPlace>(platform::CPUPlace dst_place,
-                                                  void* dst,
-                                                  platform::GPUPlace src_place,
-                                                  const void* src, size_t num) {
-  platform::SetDeviceId(src_place.device);
-  platform::GpuMemcpySync(dst, src, num, cudaMemcpyDeviceToHost);
-}
-
-template <>
-void Copy<platform::GPUPlace, platform::CPUPlace>(platform::GPUPlace dst_place,
-                                                  void* dst,
-                                                  platform::CPUPlace src_place,
-                                                  const void* src, size_t num) {
-  platform::SetDeviceId(dst_place.device);
-  platform::GpuMemcpySync(dst, src, num, cudaMemcpyHostToDevice);
-}
-
-template <>
-void Copy<platform::GPUPlace, platform::GPUPlace>(platform::GPUPlace dst_place,
-                                                  void* dst,
-                                                  platform::GPUPlace src_place,
-                                                  const void* src, size_t num) {
-  platform::SetDeviceId(dst_place.device);
-  platform::GpuMemcpySync(dst, src, num, cudaMemcpyDeviceToDevice);
-}
-
 #endif
 
 }  // namespace memory

paddle/memory/memory.cc

@@ -83,12 +83,12 @@ BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) {
 }
 
 template <>
-size_t Used<platform::GPUPlace>(platform::GPUPlace place) {
+size_t Used<platform::CUDAPlace>(platform::CUDAPlace place) {
   return GetGPUBuddyAllocator(place.device)->Used();
 }
 
 template <>
-void* Alloc<platform::GPUPlace>(platform::GPUPlace place, size_t size) {
+void* Alloc<platform::CUDAPlace>(platform::CUDAPlace place, size_t size) {
   auto* buddy_allocator = GetGPUBuddyAllocator(place.device);
   auto* ptr = buddy_allocator->Alloc(size);
   if (ptr == nullptr) {
@@ -101,14 +101,14 @@ void* Alloc<platform::GPUPlace>(platform::GPUPlace place, size_t size) {
     LOG(WARNING) << "total " << total;
     LOG(WARNING) << "GpuMinChunkSize " << platform::GpuMinChunkSize();
     LOG(WARNING) << "GpuMaxChunkSize " << platform::GpuMaxChunkSize();
-    LOG(WARNING) << "GPU memory used: " << Used<platform::GPUPlace>(place);
+    LOG(WARNING) << "GPU memory used: " << Used<platform::CUDAPlace>(place);
     platform::SetDeviceId(cur_dev);
   }
   return ptr;
 }
 
 template <>
-void Free<platform::GPUPlace>(platform::GPUPlace place, void* p) {
+void Free<platform::CUDAPlace>(platform::CUDAPlace place, void* p) {
   GetGPUBuddyAllocator(place.device)->Free(p);
 }
 

paddle/memory/memory_test.cc

@@ -82,7 +82,7 @@ TEST(BuddyAllocator, CPUMultAlloc) {
 
 #ifdef PADDLE_WITH_CUDA
 
-size_t align(size_t size, paddle::platform::GPUPlace place) {
+size_t align(size_t size, paddle::platform::CUDAPlace place) {
   size += sizeof(paddle::memory::detail::Metadata);
   size_t alignment = paddle::platform::GpuMinChunkSize();
   size_t remaining = size % alignment;
@@ -94,7 +94,7 @@ TEST(BuddyAllocator, GPUAllocation) {
 
   EXPECT_EQ(p, nullptr);
 
-  paddle::platform::GPUPlace gpu(0);
+  paddle::platform::CUDAPlace gpu(0);
   p = paddle::memory::Alloc(gpu, 4096);
 
   EXPECT_NE(p, nullptr);
@@ -103,7 +103,7 @@ TEST(BuddyAllocator, GPUAllocation) {
 }
 
 TEST(BuddyAllocator, GPUMultAlloc) {
-  paddle::platform::GPUPlace gpu;
+  paddle::platform::CUDAPlace gpu;
 
   std::unordered_map<void *, size_t> ps;
 

paddle/operators/accuracy_op.cc

@@ -53,7 +53,7 @@ class AccuracyOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("Out")->type()),

paddle/operators/accuracy_op.cu

@@ -56,7 +56,7 @@ class AccuracyOpCUDAKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
+                   "It must use CUDAPlace.");
     auto* inference = ctx.Input<Tensor>("Out");
     auto* indices = ctx.Input<Tensor>("Indices");
     auto* label = ctx.Input<Tensor>("Label");

paddle/operators/adam_op.h

@@ -13,59 +13,113 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
-#include "paddle/framework/eigen.h"
+#include <math.h>  // for sqrt in CPU and CUDA
 #include "paddle/framework/op_registry.h"
+#include "paddle/operators/detail/safe_ref.h"
+#include "paddle/platform/for_range.h"
 
 namespace paddle {
 namespace operators {
 
+template <typename T>
+struct AdamFunctor {
+  T beta1_;
+  T beta2_;
+  T epsilon_;
+
+  const T* beta1_pow_;
+  const T* beta2_pow_;
+  const T* moment1_;
+  T* moment1_out_;
+  const T* moment2_;
+  T* moment2_out_;
+  const T* lr_;
+  const T* grad_;
+  const T* param_;
+  T* param_out_;
+
+  AdamFunctor(T beta1, T beta2, T epsilon, const T* beta1_pow,
+              const T* beta2_pow, const T* mom1, T* mom1_out, const T* mom2,
+              T* mom2_out, const T* lr, const T* grad, const T* param,
+              T* param_out)
+      : beta1_(beta1),
+        beta2_(beta2),
+        epsilon_(epsilon),
+        beta1_pow_(beta1_pow),
+        beta2_pow_(beta2_pow),
+        moment1_(mom1),
+        moment1_out_(mom1_out),
+        moment2_(mom2),
+        moment2_out_(mom2_out),
+        lr_(lr),
+        grad_(grad),
+        param_(param),
+        param_out_(param_out) {}
+
+  inline HOSTDEVICE void operator()(size_t i) const {
+    // Merge all memory access together.
+    T g = grad_[i];
+    T mom1 = moment1_[i];
+    T mom2 = moment2_[i];
+    T lr = *lr_;
+    T beta1_pow = *beta1_pow_;
+    T beta2_pow = *beta2_pow_;
+    T p = param_[i];
+
+    // Calculation
+    lr *= sqrt(1 - beta2_pow) / (1 - beta1_pow);
+    mom1 = beta1_ * mom1 + (1 - beta1_) * g;
+    mom2 = beta2_ * mom2 + (1 - beta2_) * g * g;
+    p -= lr * (mom1 / (sqrt(mom2) + epsilon_));
+
+    // Write back to global memory
+    moment1_out_[i] = mom1;
+    moment2_out_[i] = mom2;
+    param_out_[i] = p;
+  }
+};
+
 template <typename DeviceContext, typename T>
 class AdamOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    auto param_out_tensor = ctx.Output<framework::Tensor>("ParamOut");
-    auto moment1_out_tensor = ctx.Output<framework::Tensor>("Moment1Out");
-    auto moment2_out_tensor = ctx.Output<framework::Tensor>("Moment2Out");
-
-    param_out_tensor->mutable_data<T>(ctx.GetPlace());
-    moment1_out_tensor->mutable_data<T>(ctx.GetPlace());
-    moment2_out_tensor->mutable_data<T>(ctx.GetPlace());
+    using paddle::framework::LoDTensor;
+    using paddle::operators::detail::Ref;
 
     T beta1 = static_cast<T>(ctx.Attr<float>("beta1"));
     T beta2 = static_cast<T>(ctx.Attr<float>("beta2"));
     T epsilon = static_cast<T>(ctx.Attr<float>("epsilon"));
+    auto& param = Ref(ctx.Input<LoDTensor>("Param"), "Must set Param");
+    auto& grad = Ref(ctx.Input<LoDTensor>("Grad"), "Must set Grad");
+    auto& mom1 = Ref(ctx.Input<LoDTensor>("Moment1"), "Must set Moment1");
+    auto& mom2 = Ref(ctx.Input<LoDTensor>("Moment2"), "Must set Moment2");
+    auto& lr =
+        Ref(ctx.Input<LoDTensor>("LearningRate"), "Must set LearningRate");
+
+    auto& beta1_pow =
+        Ref(ctx.Input<LoDTensor>("Beta1Pow"), "Must set Beta1Pow");
+    auto& beta2_pow =
+        Ref(ctx.Input<LoDTensor>("Beta2Pow"), "Must set Beta2Pow");
+
+    auto& param_out =
+        Ref(ctx.Output<LoDTensor>("ParamOut"), "Must set ParamOut");
+    auto& mom1_out =
+        Ref(ctx.Output<LoDTensor>("Moment1Out"), "Must set Moment1Out");
+    auto& mom2_out =
+        Ref(ctx.Output<LoDTensor>("Moment2Out"), "Must set Moment1Out");
 
-    auto param = framework::EigenVector<T>::Flatten(
-        *ctx.Input<framework::Tensor>("Param"));
-    auto grad = framework::EigenVector<T>::Flatten(
-        *ctx.Input<framework::Tensor>("Grad"));
-    auto moment1 = framework::EigenVector<T>::Flatten(
-        *ctx.Input<framework::Tensor>("Moment1"));
-    auto moment2 = framework::EigenVector<T>::Flatten(
-        *ctx.Input<framework::Tensor>("Moment2"));
-    auto lr = framework::EigenVector<T>::Flatten(
-        *ctx.Input<framework::Tensor>("LearningRate"));
-    auto beta1_pow = framework::EigenVector<T>::Flatten(
-        *ctx.Input<framework::Tensor>("Beta1Pow"));
-    auto beta2_pow = framework::EigenVector<T>::Flatten(
-        *ctx.Input<framework::Tensor>("Beta2Pow"));
-    auto param_out = framework::EigenVector<T>::Flatten(*param_out_tensor);
-    auto moment1_out = framework::EigenVector<T>::Flatten(*moment1_out_tensor);
-    auto moment2_out = framework::EigenVector<T>::Flatten(*moment2_out_tensor);
-    auto* place = ctx.template device_context<DeviceContext>().eigen_device();
-
-    moment1_out.device(*place) = beta1 * moment1 + (1 - beta1) * grad;
-    moment2_out.device(*place) = beta2 * moment2 + (1 - beta2) * grad.square();
-
-    // All of these are tensors of 1 element
-    auto lr_t = lr * (1 - beta2_pow).sqrt() / (1 - beta1_pow);
-    // Eigen does not support automatic broadcast
-    // Get dimensions of moment vector to broadcast lr_t
-    Eigen::DSizes<int, 1> m_dsize(moment1_out_tensor->numel());
-    param_out.device(*place) =
-        param -
-        lr_t.broadcast(m_dsize) *
-            (moment1_out / (moment2_out.sqrt() + epsilon));
+    AdamFunctor<T> functor(beta1, beta2, epsilon, beta1_pow.template data<T>(),
+                           beta2_pow.template data<T>(),
+                           mom1.template data<T>(),
+                           mom1_out.template mutable_data<T>(ctx.GetPlace()),
+                           mom2.template data<T>(),
+                           mom2_out.template mutable_data<T>(ctx.GetPlace()),
+                           lr.template data<T>(), grad.template data<T>(),
+                           param.template data<T>(),
+                           param_out.template mutable_data<T>(ctx.GetPlace()));
+    platform::ForRange<DeviceContext> for_range(
+        static_cast<const DeviceContext&>(ctx.device_context()), param.numel());
+    for_range(functor);
   }
 };
 

paddle/operators/array_operator.h

@@ -15,6 +15,7 @@
 #pragma once
 #include "paddle/framework/lod_tensor_array.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/platform/device_context.h"
 
 namespace paddle {
 namespace operators {
@@ -27,11 +28,16 @@ class ArrayOp : public framework::OperatorBase {
 
  protected:
   size_t GetOffset(const framework::Scope &scope,
-                   const platform::DeviceContext &dev_ctx) const {
+                   const platform::Place &place) const {
     auto *i = scope.FindVar(Input("I"));
     PADDLE_ENFORCE(i != nullptr, "I must be set");
     auto &i_tensor = i->Get<framework::LoDTensor>();
     PADDLE_ENFORCE_EQ(i_tensor.numel(), 1);
+
+    // get device context from pool
+    platform::DeviceContextPool &pool = platform::DeviceContextPool::Get();
+    auto &dev_ctx = *pool.Borrow(place);
+
     size_t offset;
     if (platform::is_gpu_place(i_tensor.place())) {
       // FIXME: Avoid copy from GPU to CPU

paddle/operators/array_to_lod_tensor_op.cc

@@ -12,10 +12,12 @@
    See the License for the specific language governing permissions and
    limitations under the License. */
 #include <numeric>
+
 #include "paddle/framework/lod_rank_table.h"
 #include "paddle/framework/lod_tensor_array.h"
 #include "paddle/framework/op_registry.h"
 #include "paddle/memory/memcpy.h"
+#include "paddle/platform/device_context.h"
 
 namespace paddle {
 namespace operators {
@@ -30,7 +32,7 @@ class ArrayToLoDTensorOp : public framework::OperatorBase {
                      const framework::AttributeMap &attrs)
       : OperatorBase(type, inputs, outputs, attrs) {}
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &dev_place) const override {
     auto &x = scope.FindVar(Input("X"))->Get<framework::LoDTensorArray>();
     auto &rank_table =
         scope.FindVar(Input("RankTable"))->Get<framework::LoDRankTable>();
@@ -103,6 +105,10 @@ class ArrayToLoDTensorOp : public framework::OperatorBase {
           continue;
         }
         auto slice = out->Slice(out_offset, out_offset + len);
+
+        platform::DeviceContextPool &pool = platform::DeviceContextPool::Get();
+        auto &dev_ctx = *pool.Borrow(place);
+
         framework::CopyFrom(x[x_idx].Slice(start_offset, end_offset), place,
                             dev_ctx, &slice);
         out_offset += len;

paddle/operators/assign_op.cc

@@ -15,6 +15,7 @@
 #include "paddle/framework/data_type.h"
 #include "paddle/framework/op_registry.h"
 #include "paddle/framework/var_type.h"
+#include "paddle/platform/device_context.h"
 
 namespace paddle {
 namespace operators {
@@ -71,7 +72,7 @@ class AssignOp : public framework::OperatorBase {
            const framework::AttributeMap &attrs)
       : OperatorBase(type, inputs, outputs, attrs) {}
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &place) const override {
     auto *x = scope.FindVar(Input("X"));
     if (x == nullptr) {
       return;
@@ -80,6 +81,10 @@ class AssignOp : public framework::OperatorBase {
     PADDLE_ENFORCE(
         out != nullptr,
         "The Output(Out) should not be null if the Input(X) is set.");
+
+    platform::DeviceContextPool &pool = platform::DeviceContextPool::Get();
+    auto &dev_ctx = *pool.Borrow(place);
+
     framework::VisitVarType(*x, AssignFunctor(out, dev_ctx));
   }
 };

paddle/operators/auc_op.cc

@@ -39,7 +39,7 @@ class AucOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("Out")->type()),

paddle/operators/batch_norm_op.cc

@@ -13,12 +13,14 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/operators/batch_norm_op.h"
+#include "paddle/framework/data_layout.h"
 
 namespace paddle {
 namespace operators {
 
 using Tensor = framework::Tensor;
 using LoDTensor = framework::LoDTensor;
+using DataLayout = framework::DataLayout;
 
 template <typename T>
 using EigenArrayMap =
@@ -60,14 +62,14 @@ class BatchNormOp : public framework::OperatorWithKernel {
                       "Variance and VarianceOut should share the same memory");
 
     const auto x_dims = ctx->GetInputDim("X");
-    const TensorFormat tensor_format =
-        StringToTensorFormat(ctx->Attrs().Get<std::string>("tensor_format"));
+    const DataLayout data_layout = framework::StringToDataLayout(
+        ctx->Attrs().Get<std::string>("data_layout"));
 
     PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5,
                    "Input X must have 2 to 5 dimensions.");
 
     const int C =
-        (tensor_format == TensorFormat::NCHW ? x_dims[1]
+        (data_layout == DataLayout::kNCHW ? x_dims[1]
                                           : x_dims[x_dims.size() - 1]);
 
     PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale").size(), 1UL);
@@ -90,7 +92,7 @@ class BatchNormOpMaker : public framework::OpProtoAndCheckerMaker {
     AddAttr<bool>("is_test", "").SetDefault(false);
     AddAttr<float>("momentum", "").SetDefault(0.9);
     AddAttr<float>("epsilon", "").SetDefault(1e-5);
-    AddAttr<std::string>("tensor_format", "").SetDefault("NCHW");
+    AddAttr<std::string>("data_layout", "").SetDefault("NCHW");
     AddInput("X", "The input tensor");
     AddInput("Scale",
              "Scale is a 1-dimensional tensor of size C "
@@ -141,9 +143,9 @@ class BatchNormKernel<platform::CPUDeviceContext, T>
     const float epsilon = ctx.Attr<float>("epsilon");
     const float momentum = ctx.Attr<float>("momentum");
     const bool is_test = ctx.Attr<bool>("is_test");
-    const std::string tensor_format_str =
-        ctx.Attr<std::string>("tensor_format");
-    const TensorFormat tensor_format = StringToTensorFormat(tensor_format_str);
+    const std::string data_layout_str = ctx.Attr<std::string>("data_layout");
+    const DataLayout data_layout =
+        framework::StringToDataLayout(data_layout_str);
 
     const auto *x = ctx.Input<Tensor>("X");
     const auto &x_dims = x->dims();
@@ -151,7 +153,7 @@ class BatchNormKernel<platform::CPUDeviceContext, T>
                    "The Input dim size should be between 2 and 5");
     const int N = x_dims[0];
     const int C =
-        (tensor_format == TensorFormat::NCHW ? x_dims[1]
+        (data_layout == DataLayout::kNCHW ? x_dims[1]
                                           : x_dims[x_dims.size() - 1]);
     const int sample_size = x->numel() / N / C;
 
@@ -177,8 +179,8 @@ class BatchNormKernel<platform::CPUDeviceContext, T>
       saved_mean_e.setZero();
       saved_variance_e.setZero();
 
-      switch (tensor_format) {
-        case TensorFormat::NCHW: {
+      switch (data_layout) {
+        case DataLayout::kNCHW: {
           ConstEigenArrayMap<T> x_arr(x->data<T>(), sample_size, N * C);
           for (int nc = 0; nc < N * C; ++nc) {
             saved_mean_e(nc % C) += x_arr.col(nc).sum();
@@ -191,7 +193,7 @@ class BatchNormKernel<platform::CPUDeviceContext, T>
           saved_variance_e /= N * sample_size;
           break;
         }
-        case TensorFormat::NHWC: {
+        case DataLayout::kNHWC: {
           ConstEigenArrayMap<T> x_arr(x->data<T>(), C, N * sample_size);
           for (int i = 0; i < N * sample_size; ++i) {
             saved_mean_e += x_arr.col(i);
@@ -205,7 +207,7 @@ class BatchNormKernel<platform::CPUDeviceContext, T>
           break;
         }
         default:
-          PADDLE_THROW("Unknown storage order: %s", tensor_format_str);
+          PADDLE_THROW("Unknown storage order: %s", data_layout_str);
       }
 
       EigenVectorArrayMap<T> running_mean_arr(
@@ -247,8 +249,8 @@ class BatchNormKernel<platform::CPUDeviceContext, T>
     Eigen::Array<T, Eigen::Dynamic, 1> new_bias =
         bias_arr - mean_arr * inv_std * scale_arr;
 
-    switch (tensor_format) {
-      case TensorFormat::NCHW: {
+    switch (data_layout) {
+      case DataLayout::kNCHW: {
         EigenArrayMap<T> y_arr(y->mutable_data<T>(ctx.GetPlace()), sample_size,
                                N * C);
         ConstEigenArrayMap<T> x_arr(x->data<T>(), sample_size, N * C);
@@ -257,7 +259,7 @@ class BatchNormKernel<platform::CPUDeviceContext, T>
         }
         break;
       }
-      case TensorFormat::NHWC: {
+      case DataLayout::kNHWC: {
         EigenArrayMap<T>(y->mutable_data<T>(ctx.GetPlace()), C,
                          N * sample_size) =
             (ConstEigenArrayMap<T>(x->data<T>(), C, N * sample_size).colwise() *
@@ -267,7 +269,7 @@ class BatchNormKernel<platform::CPUDeviceContext, T>
         break;
       }
       default:
-        PADDLE_THROW("Unknown storage order: %d", tensor_format);
+        PADDLE_THROW("Unknown storage order: %d", data_layout);
     }
   }
 };
@@ -290,10 +292,10 @@ class BatchNormGradOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("Bias")), "");
 
     const auto x_dims = ctx->GetInputDim("X");
-    const TensorFormat tensor_format =
-        StringToTensorFormat(ctx->Attrs().Get<std::string>("tensor_format"));
+    const DataLayout data_layout = framework::StringToDataLayout(
+        ctx->Attrs().Get<std::string>("data_layout"));
     const int C =
-        (tensor_format == TensorFormat::NCHW ? x_dims[1]
+        (data_layout == DataLayout::kNCHW ? x_dims[1]
                                           : x_dims[x_dims.size() - 1]);
 
     ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
@@ -302,7 +304,7 @@ class BatchNormGradOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
     const auto *var = ctx.InputVar(framework::GradVarName("Y"));
     if (var == nullptr) {
@@ -333,9 +335,9 @@ class BatchNormGradKernel<platform::CPUDeviceContext, T>
     const auto *saved_mean = ctx.Input<Tensor>("SavedMean");
     // SavedVariance have been reverted in forward operator
     const auto *saved_inv_variance = ctx.Input<Tensor>("SavedVariance");
-    const std::string tensor_format_str =
-        ctx.Attr<std::string>("tensor_format");
-    const TensorFormat tensor_format = StringToTensorFormat(tensor_format_str);
+    const std::string data_layout_str = ctx.Attr<std::string>("data_layout");
+    const DataLayout data_layout =
+        framework::StringToDataLayout(data_layout_str);
 
     // Get the size for each dimension.
     // NCHW [batch_size, in_channels, in_height, in_width]
@@ -344,7 +346,7 @@ class BatchNormGradKernel<platform::CPUDeviceContext, T>
                    "The Input dim size should be between 2 and 5");
     const int N = x_dims[0];
     const int C =
-        (tensor_format == TensorFormat::NCHW ? x_dims[1]
+        (data_layout == DataLayout::kNCHW ? x_dims[1]
                                           : x_dims[x_dims.size() - 1]);
     const int sample_size = x->numel() / N / C;
 
@@ -376,8 +378,8 @@ class BatchNormGradKernel<platform::CPUDeviceContext, T>
 
     const auto scale_inv_var_nhw = scale_arr * inv_var_arr / (N * sample_size);
 
-    switch (tensor_format) {
-      case TensorFormat::NCHW: {
+    switch (data_layout) {
+      case DataLayout::kNCHW: {
         ConstEigenArrayMap<T> x_arr(x->data<T>(), sample_size, N * C);
         ConstEigenArrayMap<T> d_y_arr(d_y->data<T>(), sample_size, N * C);
         EigenArrayMap<T> d_x_arr(d_x->mutable_data<T>(ctx.GetPlace()),
@@ -400,7 +402,7 @@ class BatchNormGradKernel<platform::CPUDeviceContext, T>
         }
         break;
       }
-      case TensorFormat::NHWC: {
+      case DataLayout::kNHWC: {
         ConstEigenArrayMap<T> x_arr(x->data<T>(), C, N * sample_size);
         ConstEigenArrayMap<T> d_y_arr(d_y->data<T>(), C, N * sample_size);
         EigenArrayMap<T> d_x_arr(d_x->mutable_data<T>(ctx.GetPlace()), C,
@@ -425,7 +427,7 @@ class BatchNormGradKernel<platform::CPUDeviceContext, T>
         break;
       }
       default:
-        PADDLE_THROW("Unknown storage order: %s", tensor_format_str);
+        PADDLE_THROW("Unknown storage order: %s", data_layout_str);
     }
   }
 };

paddle/operators/batch_norm_op.cu.cc

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/operators/batch_norm_op.h"
+#include "paddle/framework/data_layout.h"
 
 #include <cfloat>
 #include "paddle/operators/math/math_function.h"
@@ -22,12 +23,12 @@ namespace paddle {
 namespace operators {
 
 using Tensor = framework::Tensor;
+using DataLayout = framework::DataLayout;
 template <typename T>
 using CudnnDataType = platform::CudnnDataType<T>;
 
-void ExtractNCWHD(const framework::DDim &dims,
-                  const TensorFormat &tensor_format, int *N, int *C, int *H,
-                  int *W, int *D) {
+void ExtractNCWHD(const framework::DDim &dims, const DataLayout &data_layout,
+                  int *N, int *C, int *H, int *W, int *D) {
   *N = dims[0];
   if (dims.size() == 2) {
     *C = dims[1];
@@ -35,13 +36,13 @@ void ExtractNCWHD(const framework::DDim &dims,
     *W = 1;
     *D = 1;
   } else {
-    *C = tensor_format == TensorFormat::NCHW ? dims[1] : dims[dims.size() - 1];
-    *H = tensor_format == TensorFormat::NCHW ? dims[2] : dims[1];
+    *C = data_layout == DataLayout::kNCHW ? dims[1] : dims[dims.size() - 1];
+    *H = data_layout == DataLayout::kNCHW ? dims[2] : dims[1];
     *W = dims.size() > 3
-             ? (tensor_format == TensorFormat::NCHW ? dims[3] : dims[2])
+             ? (data_layout == DataLayout::kNCHW ? dims[3] : dims[2])
              : 1;
     *D = dims.size() > 4
-             ? (tensor_format == TensorFormat::NCHW ? dims[4] : dims[3])
+             ? (data_layout == DataLayout::kNCHW ? dims[4] : dims[3])
              : 1;
   }
 }
@@ -52,13 +53,13 @@ class BatchNormKernel<platform::CUDADeviceContext, T>
  public:
   void Compute(const framework::ExecutionContext &ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
+                   "It must use CUDAPlace.");
     double epsilon = static_cast<double>(ctx.Attr<float>("epsilon"));
     const float momentum = ctx.Attr<float>("momentum");
     const bool is_test = ctx.Attr<bool>("is_test");
-    const std::string tensor_format_str =
-        ctx.Attr<std::string>("tensor_format");
-    const TensorFormat tensor_format = StringToTensorFormat(tensor_format_str);
+    const std::string data_layout_str = ctx.Attr<std::string>("data_layout");
+    const DataLayout data_layout =
+        framework::StringToDataLayout(data_layout_str);
 
     // Get the size for each dimension.
     // NCHW [batch_size, in_channels, in_height, in_width]
@@ -67,7 +68,7 @@ class BatchNormKernel<platform::CUDADeviceContext, T>
     PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5,
                    "The Input dim size should be between 2 and 5");
     int N, C, H, W, D;
-    ExtractNCWHD(x_dims, tensor_format, &N, &C, &H, &W, &D);
+    ExtractNCWHD(x_dims, data_layout, &N, &C, &H, &W, &D);
 
     // ------------------- cudnn descriptors ---------------------
     cudnnTensorDescriptor_t data_desc_;
@@ -93,7 +94,7 @@ class BatchNormKernel<platform::CUDADeviceContext, T>
     VLOG(1) << "Setting descriptors.";
     std::vector<int> dims;
     std::vector<int> strides;
-    if (tensor_format == TensorFormat::NCHW) {
+    if (data_layout == DataLayout::kNCHW) {
       dims = {N, C, H, W, D};
       strides = {C * H * W * D, H * W * D, W * D, D, 1};
     } else {
@@ -178,11 +179,11 @@ class BatchNormGradKernel<platform::CUDADeviceContext, T>
  public:
   void Compute(const framework::ExecutionContext &ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
+                   "It must use CUDAPlace.");
     double epsilon = static_cast<double>(ctx.Attr<float>("epsilon"));
-    const std::string tensor_format_str =
-        ctx.Attr<std::string>("tensor_format");
-    const TensorFormat tensor_format = StringToTensorFormat(tensor_format_str);
+    const std::string data_layout_str = ctx.Attr<std::string>("data_layout");
+    const DataLayout data_layout =
+        framework::StringToDataLayout(data_layout_str);
     const auto *x = ctx.Input<Tensor>("X");
     const auto *d_y = ctx.Input<Tensor>(framework::GradVarName("Y"));
     const auto *scale = ctx.Input<Tensor>("Scale");
@@ -192,7 +193,7 @@ class BatchNormGradKernel<platform::CUDADeviceContext, T>
     PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5,
                    "The Input dim size should be between 2 and 5");
     int N, C, H, W, D;
-    ExtractNCWHD(x_dims, tensor_format, &N, &C, &H, &W, &D);
+    ExtractNCWHD(x_dims, data_layout, &N, &C, &H, &W, &D);
 
     PADDLE_ENFORCE_EQ(scale->dims().size(), 1UL);
     PADDLE_ENFORCE_EQ(scale->dims()[0], C);
@@ -219,7 +220,7 @@ class BatchNormGradKernel<platform::CUDADeviceContext, T>
 
     std::vector<int> dims;
     std::vector<int> strides;
-    if (tensor_format == TensorFormat::NCHW) {
+    if (data_layout == DataLayout::kNCHW) {
       dims = {N, C, H, W, D};
       strides = {C * H * W * D, H * W * D, W * D, D, 1};
     } else {

paddle/operators/batch_norm_op.h

@@ -19,21 +19,6 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-enum TensorFormat {
-  NHWC = 0,
-  NCHW = 1,
-};
-
-inline TensorFormat StringToTensorFormat(const std::string& str) {
-  if (str == "NHWC" || str == "nhwc") {
-    return TensorFormat::NHWC;
-  } else if (str == "NCHW" || str == "nchw") {
-    return TensorFormat::NCHW;
-  } else {
-    PADDLE_THROW("Unknown storage order string: %s", str);
-  }
-}
-
 template <typename DeviceContext, typename T>
 class BatchNormKernel : public framework::OpKernel<T> {
  public:

paddle/operators/beam_search_decode_op.cc

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/operators/beam_search_decode_op.h"
+#include "paddle/platform/device_context.h"
 
 namespace paddle {
 namespace operators {
@@ -55,7 +56,10 @@ class BeamSearchDecodeOp : public framework::OperatorBase {
                      const framework::AttributeMap& attrs)
       : OperatorBase(type, inputs, outputs, attrs) {}
   void Run(const framework::Scope& scope,
-           const platform::DeviceContext& dev_ctx) const override {
+           const platform::Place& dev_place) const override {
+    platform::DeviceContextPool& pool = platform::DeviceContextPool::Get();
+    auto& dev_ctx = *pool.Borrow(dev_place);
+
     framework::ExecutionContext ctx(*this, scope, dev_ctx);
 
     const LoDTensorArray* ids = ctx.Input<LoDTensorArray>("Ids");

paddle/operators/beam_search_op.h

@@ -189,7 +189,7 @@ class BeamSearchOp : public framework::OperatorBase {
   }
 
   void Run(const framework::Scope& scope,
-           const platform::DeviceContext& dev_ctx) const override {
+           const platform::Place& dev_place) const override {
     LOG(INFO) << "run beam search op";
     auto ids_var = scope.FindVar(Input("ids"));
     auto scores_var = scope.FindVar(Input("scores"));

paddle/operators/chunk_eval_op.cc

@@ -55,7 +55,7 @@ class ChunkEvalOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(framework::proto::DataType::FP32,
                                    ctx.device_context());

paddle/operators/compare_op.cc

@@ -66,9 +66,9 @@ class CompareOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
-    framework::OpKernelType kt = OperatorWithKernel::GetKernelType(ctx);
+    framework::OpKernelType kt = OperatorWithKernel::GetActualKernelType(ctx);
     // CompareOp kernel's device type is decided by input tensor place
     kt.place_ = ctx.Input<framework::LoDTensor>("X")->place();
     return kt;

paddle/operators/concat_op.cc

@@ -98,8 +98,8 @@ class ConcatOpGrad : public framework::OperatorWithKernel {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP(concat, ops::ConcatOp, ops::ConcatOpMaker, concat_grad,
-            ops::ConcatOpGrad)
+REGISTER_OP_EX(concat, ops::ConcatOp, ops::ConcatOpMaker, concat_grad,
+               ops::ConcatOpGrad, false)
 REGISTER_OP_CPU_KERNEL(concat,
                        ops::ConcatKernel<paddle::platform::CPUPlace, float>)
 REGISTER_OP_CPU_KERNEL(concat_grad,

paddle/operators/cond_op.cc

@@ -13,9 +13,9 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/operators/cond_op.h"
-
 #include "paddle/operators/gather.h"
 #include "paddle/operators/scatter.h"
+#include "paddle/platform/device_context.h"
 
 namespace paddle {
 namespace operators {
@@ -193,12 +193,15 @@ void CondOp::MergeDataFromSubnet(const framework::Scope& scope,
   }
 }
 
-void CondOp::Run(const Scope& scope,
-                 const platform::DeviceContext& dev_ctx) const {
+void CondOp::Run(const Scope& scope, const platform::Place& place) const {
+  // get device context from pool
+  platform::DeviceContextPool& pool = platform::DeviceContextPool::Get();
+  auto& dev_ctx = *pool.Borrow(place);
+
   PrepareDataForSubnet(scope, dev_ctx);
   std::vector<framework::Scope*>& sub_scopes = GetSubScopes(scope);
   for (int i = 0; i < BRANCH_NUM; ++i) {
-    sub_net_op_[i]->Run(*sub_scopes[i], dev_ctx);
+    sub_net_op_[i]->Run(*sub_scopes[i], place);
   }
   MergeDataFromSubnet(scope, dev_ctx);
 }

paddle/operators/cond_op.h

@@ -78,7 +78,7 @@ class CondOp : public framework::OperatorBase {
   }
 
   void Run(const framework::Scope& scope,
-           const platform::DeviceContext& dev_ctx) const override;
+           const platform::Place& place) const override;
 
  private:
   const int TRUE_BRANCH = 0;

paddle/operators/conditional_block_op.cc

@@ -51,7 +51,7 @@ class ConditionalBlockOp : public ConditionalOp {
                      const framework::AttributeMap &attrs)
       : ConditionalOp(type, inputs, outputs, attrs) {}
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &dev_place) const override {
     auto xs = InputTensors(scope);
     bool need_run = std::all_of(
         xs.begin(), xs.end(),
@@ -65,8 +65,8 @@ class ConditionalBlockOp : public ConditionalOp {
       scopes->front() = &scope.NewScope();
       auto &cur_scope = *scopes->front();
 
+      framework::Executor exec(dev_place);
       auto *block = Attr<framework::BlockDesc *>("sub_block");
-      framework::Executor exec(dev_ctx);
       exec.Run(*block->Program(), &cur_scope, block->ID(), false);
     }
   }
@@ -104,7 +104,7 @@ class ConditionalBlockGradOp : public ConditionalOp {
                          const framework::AttributeMap &attrs)
       : ConditionalOp(type, inputs, outputs, attrs) {}
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &dev_place) const override {
     auto xs = this->InputTensors(scope);
     bool need_run = std::all_of(
         xs.begin(), xs.end(),
@@ -116,21 +116,21 @@ class ConditionalBlockGradOp : public ConditionalOp {
       auto &scopes = scope_var->Get<std::vector<framework::Scope *>>();
       framework::Scope &cur_scope = *scopes[0];
 
+      framework::Executor exec(dev_place);
       auto *block = Attr<framework::BlockDesc *>("sub_block");
-      framework::Executor exec(dev_ctx);
       exec.Run(*block->Program(), &cur_scope, block->ID(), false);
 
-      AssignLocalGradientToGlobal(dev_ctx, cur_scope, Inputs("Params"),
+      AssignLocalGradientToGlobal(dev_place, cur_scope, Inputs("Params"),
                                   Outputs(framework::GradVarName("Params")));
 
-      AssignLocalGradientToGlobal(dev_ctx, cur_scope, Inputs("X"),
+      AssignLocalGradientToGlobal(dev_place, cur_scope, Inputs("X"),
                                   Outputs(framework::GradVarName("X")));
     }
   }
 
  private:
   void AssignLocalGradientToGlobal(
-      const platform::DeviceContext &dev_ctx, const framework::Scope &cur_scope,
+      const platform::Place &place, const framework::Scope &cur_scope,
       const std::vector<std::string> &p_names,
       const std::vector<std::string> &pg_names) const {
     for (size_t i = 0; i < p_names.size(); ++i) {
@@ -144,7 +144,7 @@ class ConditionalBlockGradOp : public ConditionalOp {
       auto assign = framework::OpRegistry::CreateOp(
           "assign", {{"X", {new_in_grad_name}}}, {{"Out", {out_grad_name}}},
           framework::AttributeMap{});
-      assign->Run(cur_scope, dev_ctx);
+      assign->Run(cur_scope, place);
       cur_scope.Rename(new_in_grad_name, in_grad_name);
     }
   }
@@ -178,8 +178,9 @@ class ConditionalBlockGradMaker : public framework::SingleGradOpDescMaker {
     grad_op->SetInput("Out", Output("Out"));
     grad_op->SetInput(framework::GradVarName("Out"), OutputGrad("Out"));
     grad_op->SetInput("Scope", Output("Scope"));
-    grad_op->SetOutput(framework::GradVarName("X"), InputGrad("X"));
-    grad_op->SetOutput(framework::GradVarName("Params"), InputGrad("Params"));
+    grad_op->SetOutput(framework::GradVarName("X"), InputGrad("X", false));
+    grad_op->SetOutput(framework::GradVarName("Params"),
+                       InputGrad("Params", false));
     grad_op->SetBlockAttr("sub_block", *this->grad_block_[0]);
     return std::unique_ptr<framework::OpDesc>(grad_op);
   }

paddle/operators/conv_cudnn_op.cu.cc

@@ -36,7 +36,7 @@ class CudnnConvOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
+                   "It must use CUDAPlace.");
     auto* input = ctx.Input<Tensor>("Input");
     auto* filter = ctx.Input<Tensor>("Filter");
     auto* output = ctx.Output<Tensor>("Output");
@@ -130,7 +130,7 @@ class CudnnConvOpKernel : public framework::OpKernel<T> {
         handle, cudnn_input_desc, cudnn_filter_desc, cudnn_conv_desc,
         cudnn_output_desc, algo, &workspace_size_in_bytes));
     // Allocate on GPU memory
-    platform::GPUPlace gpu = boost::get<platform::GPUPlace>(ctx.GetPlace());
+    platform::CUDAPlace gpu = boost::get<platform::CUDAPlace>(ctx.GetPlace());
     cudnn_workspace = paddle::memory::Alloc(gpu, workspace_size_in_bytes);
     // ------------------- cudnn conv forward ---------------------
     T alpha = 1.0f, beta = 0.0f;
@@ -151,7 +151,7 @@ class CudnnConvGradOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
+                   "It must use CUDAPlace.");
     auto input = ctx.Input<Tensor>("Input");
     auto filter = ctx.Input<Tensor>("Filter");
     auto output_grad = ctx.Input<Tensor>(framework::GradVarName("Output"));
@@ -277,7 +277,7 @@ class CudnnConvGradOpKernel : public framework::OpKernel<T> {
     // ------------------- cudnn conv workspace ---------------------
     // Already on GPU
     void* cudnn_workspace = nullptr;
-    platform::GPUPlace gpu = boost::get<platform::GPUPlace>(ctx.GetPlace());
+    platform::CUDAPlace gpu = boost::get<platform::CUDAPlace>(ctx.GetPlace());
     cudnn_workspace = paddle::memory::Alloc(gpu, workspace_size_in_bytes);
     // ------------------- cudnn conv backward data ---------------------
     T alpha = 1.0f, beta = 0.0f;

paddle/operators/conv_transpose_cudnn_op.cu.cc

@@ -35,7 +35,7 @@ class CudnnConvTransposeOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
+                   "It must use CUDAPlace.");
     auto* input = ctx.Input<Tensor>("Input");
     auto* filter = ctx.Input<Tensor>("Filter");
     auto* output = ctx.Output<Tensor>("Output");
@@ -100,7 +100,7 @@ class CudnnConvTransposeOpKernel : public framework::OpKernel<T> {
             cudnn_output_desc, algo, &workspace_size_in_bytes));
 
     // Allocate on GPU memory
-    platform::GPUPlace gpu = boost::get<platform::GPUPlace>(ctx.GetPlace());
+    platform::CUDAPlace gpu = boost::get<platform::CUDAPlace>(ctx.GetPlace());
     cudnn_workspace = paddle::memory::Alloc(gpu, workspace_size_in_bytes);
 
     // ------------------- cudnn conv transpose forward ---------------------
@@ -120,7 +120,7 @@ class CudnnConvTransposeGradOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
+                   "It must use CUDAPlace.");
     auto input = ctx.Input<Tensor>("Input");
     auto filter = ctx.Input<Tensor>("Filter");
     auto output_grad = ctx.Input<Tensor>(framework::GradVarName("Output"));
@@ -201,7 +201,7 @@ class CudnnConvTransposeGradOpKernel : public framework::OpKernel<T> {
     // ------------------- cudnn conv workspace ---------------------
     // Already on GPU
     void* cudnn_workspace = nullptr;
-    platform::GPUPlace gpu = boost::get<platform::GPUPlace>(ctx.GetPlace());
+    platform::CUDAPlace gpu = boost::get<platform::CUDAPlace>(ctx.GetPlace());
     cudnn_workspace = paddle::memory::Alloc(gpu, workspace_size_in_bytes);
     // ------------------- cudnn conv backward data ---------------------
     // FIXME(typhoonzero): template type T may not be the same as cudnn call.

paddle/operators/crf_decoding_op.cc

@@ -120,12 +120,18 @@ class CRFDecodingOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<LoDTensor>("Emission")->type()),
         ctx.device_context());
   }
+
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::OpKernelType& actual_kernel_type) const override {
+    return framework::OpKernelType(actual_kernel_type.data_type_,
+                                   platform::CPUPlace());
+  }
 };
 }  // namespace operators
 }  // namespace paddle

paddle/operators/cross_entropy_op.cc

@@ -51,7 +51,7 @@ class CrossEntropyOp : public framework::OperatorWithKernel {
  protected:
   // Explicitly set that the data type of computation kernel of cross_entropy
   // is determined by its input "X".
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("X")->type()),
@@ -101,7 +101,7 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel {
  protected:
   // Explicitly set that the data type of computation kernel of cross_entropy
   // is determined by its input "X".
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("X")->type()),

paddle/operators/detail/recv_impl.cc

@@ -20,25 +20,57 @@ namespace detail {
 
 Status SendRecvServerImpl::SendVariable(ServerContext *context,
                                         const VariableMessage *in_var,
-                                        VariableMessage *out_var) {
-  framework::LoDTensor t;
-  // TODO(typhoonzero): desirealize in_tensor and run pserver network.
+                                        VoidMessage *out_var) {
+  // TODO(typhoonzero): support different variable types.
   std::istringstream iss(in_var->serialized());
+  framework::LoDTensor t;
   framework::DeserializeFromStream(iss, &t);
-  lodtensor_queue_.Push(std::move(t));
-  // Block util the sub graph is done.
-  t = lodtensor_return_queue_.Pop();
+  TensorWithName tensor_with_name =
+      std::make_pair(in_var->varname(), std::move(t));
+
+  var_recv_queue_.Push(std::move(tensor_with_name));
+  return Status::OK;
+}
+
+Status SendRecvServerImpl::GetVariable(ServerContext *context,
+                                       const VariableMessage *in_var,
+                                       VariableMessage *out_var) {
+  std::string get_var_name = in_var->varname();
+  auto *var = scope_->FindVar(get_var_name);
+  auto tensor = var->Get<framework::LoDTensor>();
   std::ostringstream oss;
-  // FIXME(typhoonzero): get context from op.
-  framework::SerializeToStream(oss, t, platform::CPUDeviceContext());
+  framework::SerializeToStream(oss, tensor, platform::CPUDeviceContext());
+
   std::string *varname = out_var->mutable_varname();
-  *varname = in_var->varname();
+  *varname = get_var_name;
   std::string *serialized = out_var->mutable_serialized();
   *serialized = oss.str();
+  return Status::OK;
+}
 
+Status SendRecvServerImpl::Wait(ServerContext *context,
+                                const VoidMessage *in_var,
+                                VoidMessage *out_var) {
+  {
+    std::unique_lock<std::mutex> lock(this->mutex_);
+    condition_.wait(lock, [=] { return this->done_ == true; });
+  }
   return Status::OK;
 }
 
+void SendRecvServerImpl::Reset() {
+  std::lock_guard<std::mutex> lock(this->mutex_);
+  done_ = false;
+}
+
+void SendRecvServerImpl::Done() {
+  {
+    std::lock_guard<std::mutex> lock(this->mutex_);
+    done_ = true;
+  }
+  condition_.notify_all();
+}
+
 }  // namespace detail
 }  // namespace operators
 }  // namespace paddle

paddle/operators/detail/send_impl.cc

@@ -19,10 +19,10 @@ namespace operators {
 namespace detail {
 
 bool RPCClient::SendVariable(const framework::Scope& scope,
-                             const std::string& inname,
-                             const std::string& outname) {
+                             const std::string& inname) {
   ClientContext context;
-  VariableMessage msg, out_msg;
+  VariableMessage msg;
+  VoidMessage out_msg;
   // FIXME(typhoonzero): pass device context to here.
   auto ctx = platform::CPUDeviceContext();
   auto* var = scope.FindVar(inname);
@@ -37,9 +37,26 @@ bool RPCClient::SendVariable(const framework::Scope& scope,
   msg.set_serialized(oss.str());
   Status status = stub_->SendVariable(&context, msg, &out_msg);
   if (!status.ok()) {
+    LOG(ERROR) << "gRPC error: " << status.error_message();
     return false;
   }
-  std::istringstream iss(out_msg.serialized());
+  return true;
+}
+
+bool RPCClient::GetVariable(const framework::Scope& scope,
+                            const std::string& outname) {
+  ClientContext context;
+  VariableMessage call_msg, ret_msg;
+  call_msg.set_varname(outname);
+  auto ctx = platform::CPUDeviceContext();
+  Status status = stub_->GetVariable(&context, call_msg, &ret_msg);
+  if (!status.ok()) {
+    LOG(ERROR) << "gRPC error: " << status.error_message();
+    return false;
+  }
+
+  std::istringstream iss(ret_msg.serialized());
+
   framework::LoDTensor ret_tensor;
   framework::DeserializeFromStream(iss, &ret_tensor);
   auto* outvar = scope.FindVar(outname);
@@ -49,6 +66,12 @@ bool RPCClient::SendVariable(const framework::Scope& scope,
   return true;
 }
 
+void RPCClient::Wait() {
+  ClientContext context;
+  VoidMessage call_msg, ret_msg;
+  stub_->Wait(&context, call_msg, &ret_msg);
+}
+
 }  // namespace detail
 }  // namespace operators
 }  // namespace paddle

paddle/operators/detail/send_recv.proto

@@ -19,7 +19,12 @@ package sendrecv;
 service SendRecvService {
   // For parameter server round-robin like hashing, do not split tensors.
   // Send and recv only one tensor
-  rpc SendVariable(VariableMessage) returns (VariableMessage) {}
+  // TODO(typhoonzero): add streaming API
+  rpc SendVariable(VariableMessage) returns (VoidMessage) {}
+  // Argument VariableMessage for GetVariable should only contain varname.
+  rpc GetVariable(VariableMessage) returns (VariableMessage) {}
+  // wait for one execution of the program
+  rpc Wait(VoidMessage) returns (VoidMessage) {}
 }
 
 // VariableMessage is serialized paddle variable message.

paddle/operators/detail/send_recv_impl.h

@@ -20,10 +20,6 @@
 #include "paddle/framework/selected_rows.h"
 #include "paddle/operators/detail/simple_block_queue.h"
 
-// #include <grpc++/channel.h>
-// #include <grpc++/client_context.h>
-// #include <grpc++/create_channel.h>
-// #include <grpc++/security/credentials.h>
 #include "paddle/operators/detail/send_recv.grpc.pb.h"
 #include "paddle/operators/detail/send_recv.pb.h"
 
@@ -48,24 +44,32 @@ namespace paddle {
 namespace operators {
 namespace detail {
 
+typedef std::pair<std::string, framework::LoDTensor> TensorWithName;
+
 class SendRecvServerImpl final : public SendRecvService::Service {
  public:
   explicit SendRecvServerImpl() {}
 
   Status SendVariable(ServerContext *context, const VariableMessage *in_var,
+                      VoidMessage *out_var) override;
+  Status GetVariable(ServerContext *context, const VariableMessage *in_var,
                      VariableMessage *out_var) override;
+  Status Wait(ServerContext *context, const VoidMessage *in_var,
+              VoidMessage *out_var) override;
+  void Reset();
+  void Done();
+  void SetScope(framework::Scope *scope) { scope_ = scope; };
 
-  const framework::LoDTensor Get() { return this->lodtensor_queue_.Pop(); }
-
-  void Push(const framework::LoDTensor &tensor) {
-    this->lodtensor_return_queue_.Push(tensor);
-  }
+  const TensorWithName Get() { return this->var_recv_queue_.Pop(); }
 
  private:
-  SimpleBlockQueue<framework::LoDTensor> lodtensor_queue_;
-  SimpleBlockQueue<framework::LoDTensor> lodtensor_return_queue_;
-  SimpleBlockQueue<framework::SelectedRows> selected_rows_queue_;
-  SimpleBlockQueue<framework::SelectedRows> selected_rows_return_queue_;
+  // received variable from RPC, operators fetch variable from this queue.
+  SimpleBlockQueue<TensorWithName> var_recv_queue_;
+  framework::Scope *scope_;
+  // condition of the sub program
+  std::mutex mutex_;
+  bool done_;
+  std::condition_variable condition_;
 };
 
 // RPCClient is a class to send tensors to pserver sub-network
@@ -75,8 +79,9 @@ class RPCClient {
   RPCClient(std::shared_ptr<Channel> channel)
       : stub_(SendRecvService::NewStub(channel)) {}
 
-  bool SendVariable(const framework::Scope &scope, const std::string &inname,
-                    const std::string &outname);
+  bool SendVariable(const framework::Scope &scope, const std::string &inname);
+  bool GetVariable(const framework::Scope &scope, const std::string &outname);
+  void Wait();
 
  private:
   std::unique_ptr<SendRecvService::Stub> stub_;

paddle/operators/detail/strided_memcpy.h

@@ -35,7 +35,7 @@ struct StridedMemcpyFunctor<T, 1> {
       memory::Copy(cpu_place, dst, cpu_place, src, sizeof(T) * dst_dim.head);
     } else {
 #ifdef PADDLE_WITH_CUDA
-      auto& gpu_place = boost::get<platform::GPUPlace>(place);
+      auto& gpu_place = boost::get<platform::CUDAPlace>(place);
       auto& cuda_ctx =
           reinterpret_cast<const platform::CUDADeviceContext&>(dev_ctx);
       memory::Copy(gpu_place, dst, gpu_place, src, sizeof(T) * dst_dim.head,

paddle/operators/feed_op.cc

@@ -25,7 +25,7 @@ class FeedOp : public framework::OperatorBase {
          const framework::AttributeMap &attrs)
       : OperatorBase(type, inputs, outputs, attrs) {}
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &place) const override {
     auto feed_var_name = Input("X");
     auto *feed_var = scope.FindVar(feed_var_name);
 
@@ -47,7 +47,12 @@ class FeedOp : public framework::OperatorBase {
     auto &feed_list = feed_var->Get<framework::FeedFetchList>();
     auto &feed_item = feed_list.at(static_cast<size_t>(col));
     auto *out_item = out_var->GetMutable<framework::FeedFetchType>();
-    framework::CopyFrom(feed_item, dev_ctx.GetPlace(), dev_ctx, out_item);
+
+    // get device context from pool
+    platform::DeviceContextPool &pool = platform::DeviceContextPool::Get();
+    auto &dev_ctx = *pool.Borrow(place);
+
+    framework::CopyFrom(feed_item, place, dev_ctx, out_item);
     out_item->set_lod(feed_item.lod());
   }
 };

paddle/operators/fetch_op.cc

@@ -14,6 +14,7 @@
 
 #include "paddle/framework/feed_fetch_type.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/platform/device_context.h"
 
 namespace paddle {
 namespace operators {
@@ -26,7 +27,7 @@ class FetchOp : public framework::OperatorBase {
       : OperatorBase(type, inputs, outputs, attrs) {}
 
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &place) const override {
     auto fetch_var_name = Input("X");
     auto *fetch_var = scope.FindVar(fetch_var_name);
     PADDLE_ENFORCE(fetch_var != nullptr,
@@ -51,6 +52,9 @@ class FetchOp : public framework::OperatorBase {
 
     // FIXME(yuyang18): Should we assume the fetch operator always generate
     // CPU outputs?
+    platform::DeviceContextPool &pool = platform::DeviceContextPool::Get();
+    auto &dev_ctx = *pool.Borrow(place);
+
     CopyFrom(src_item, platform::CPUPlace(), dev_ctx, &dst_item);
     dev_ctx.Wait();
     dst_item.set_lod(src_item.lod());

paddle/operators/fill_constant_batch_size_like_op.cc

@@ -49,7 +49,7 @@ class FillConstantBatchSizeLikeOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         static_cast<framework::proto::DataType>(ctx.Attr<int>("dtype")),

paddle/operators/fill_constant_op.cc

@@ -15,6 +15,7 @@ limitations under the License. */
 #include "paddle/framework/data_type.h"
 #include "paddle/framework/op_registry.h"
 #include "paddle/operators/math/math_function.h"
+#include "paddle/platform/device_context.h"
 
 namespace paddle {
 namespace operators {
@@ -33,7 +34,7 @@ class FillConstantOp : public framework::OperatorBase {
  public:
   using framework::OperatorBase::OperatorBase;
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &dev_place) const override {
     auto data_type =
         static_cast<framework::proto::DataType>(Attr<int>("dtype"));
     auto value = Attr<float>("value");
@@ -45,8 +46,11 @@ class FillConstantOp : public framework::OperatorBase {
       auto cpu = platform::CPUPlace();
       out.mutable_data(cpu, framework::ToTypeIndex(data_type));
     } else {
-      out.mutable_data(dev_ctx.GetPlace(), framework::ToTypeIndex(data_type));
+      out.mutable_data(dev_place, framework::ToTypeIndex(data_type));
     }
+
+    platform::DeviceContextPool &pool = platform::DeviceContextPool::Get();
+    auto &dev_ctx = *pool.Borrow(dev_place);
     math::set_constant(dev_ctx, &out, value);
   }
 };

paddle/operators/fill_op.cc

@@ -15,6 +15,7 @@
 #include "paddle/framework/data_type.h"
 #include "paddle/framework/op_registry.h"
 #include "paddle/operators/detail/safe_ref.h"
+#include "paddle/platform/device_context.h"
 
 namespace paddle {
 namespace operators {
@@ -42,7 +43,7 @@ class FillOp : public framework::OperatorBase {
          const framework::AttributeMap &attrs)
       : OperatorBase(type, inputs, outputs, attrs) {}
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &place) const override {
     auto &out =
         detail::Ref(detail::Ref(scope.FindVar(Output("Out")),
                                 "Cannot find variable %s", Output("Out"))
@@ -51,12 +52,11 @@ class FillOp : public framework::OperatorBase {
     auto dtype = static_cast<framework::proto::DataType>(Attr<int>("dtype"));
     platform::CPUPlace cpu;
     auto force_cpu = Attr<bool>("force_cpu");
-    out.mutable_data(force_cpu ? cpu : dev_ctx.GetPlace(),
-                     framework::ToTypeIndex(dtype));
+    out.mutable_data(force_cpu ? cpu : place, framework::ToTypeIndex(dtype));
 
     framework::LoDTensor tensor;
 
-    if (force_cpu || platform::is_cpu_place(dev_ctx.GetPlace())) {
+    if (force_cpu || platform::is_cpu_place(place)) {
       tensor.ShareDataWith(out);
     } else {
       // Always make tensor in CPU memory.
@@ -67,9 +67,11 @@ class FillOp : public framework::OperatorBase {
     framework::VisitDataType(
         dtype, FillOpVisitor(&tensor, Attr<std::vector<float>>("value")));
 
-    if (!force_cpu && platform::is_gpu_place(dev_ctx.GetPlace())) {
+    if (!force_cpu && platform::is_gpu_place(place)) {
       // Copy tensor to out
-      framework::CopyFrom(tensor, dev_ctx.GetPlace(), dev_ctx, &out);
+      platform::DeviceContextPool &pool = platform::DeviceContextPool::Get();
+      auto &dev_ctx = *pool.Borrow(place);
+      framework::CopyFrom(tensor, place, dev_ctx, &out);
     }
   }
 };

paddle/operators/gather_op.cc

@@ -40,7 +40,7 @@ class GatherOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("X")->type()),
@@ -57,7 +57,7 @@ class GatherGradOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("X")->type()),

paddle/operators/gaussian_random_op.cc

@@ -57,7 +57,7 @@ class GaussianRandomOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         static_cast<framework::proto::DataType>(ctx.Attr<int>("dtype")),

paddle/operators/increment_op.cc

@@ -52,7 +52,7 @@ class IncrementOp : public framework::OperatorBase {
       : OperatorBase(type, inputs, outputs, attrs) {}
 
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &place) const override {
     auto &x = scope.FindVar(Input("X"))->Get<framework::LoDTensor>();
     auto &out =
         *scope.FindVar(Output("Out"))->GetMutable<framework::LoDTensor>();

paddle/operators/is_empty_op.cc

@@ -29,7 +29,7 @@ class IsEmptyOp : public framework::OperatorBase {
       : OperatorBase(type, inputs, outputs, attrs) {}
 
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &place) const override {
     // get input
     auto *var = scope.FindVar(Input(kInput));
     PADDLE_ENFORCE_NOT_NULL(var);

paddle/operators/linear_chain_crf_op.cc

@@ -183,7 +183,7 @@ class LinearChainCRFOp : public framework::OperatorWithKernel {
  protected:
   // Explicitly set that the data type of computation kernel of linear_chain_crf
   // is determined by its input "Emission".
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<LoDTensor>("Emission")->type()),
@@ -242,7 +242,7 @@ class LinearChainCRFGradOp : public framework::OperatorWithKernel {
  protected:
   // Explicitly set that the data type of output of the linear_chain_crf_grad
   // operator is determined by its input: gradients of LogLikelihood.
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(

paddle/operators/linear_chain_crf_op.h

@@ -219,8 +219,8 @@ class LinearChainCRFOpKernel : public framework::OpKernel<T> {
     // operators runs on GPU device.
     auto copyTensor = [](const platform::DeviceContext& ctx, const Tensor& src,
                          Tensor* dst) {
-      dst->mutable_data<T>(platform::GPUPlace());
-      framework::CopyFrom(src, platform::GPUPlace(), ctx, dst);
+      dst->mutable_data<T>(platform::CUDAPlace());
+      framework::CopyFrom(src, platform::CUDAPlace(), ctx, dst);
     };
     copyTensor(ctx, emission_exps_src, emission_exps_dst);
     copyTensor(ctx, transition_exps_src, transition_exps_dst);
@@ -433,8 +433,8 @@ class LinearChainCRFGradOpKernel : public framework::OpKernel<T> {
     auto copyTensor = [](const platform::DeviceContext& ctx, const Tensor* src,
                          Tensor* dst) {
       if (src && dst) {
-        dst->mutable_data<T>(platform::GPUPlace());
-        framework::CopyFrom(*src, platform::GPUPlace(), ctx, dst);
+        dst->mutable_data<T>(platform::CUDAPlace());
+        framework::CopyFrom(*src, platform::CUDAPlace(), ctx, dst);
       }
     };
     copyTensor(ctx, emission_grad_src, emission_grad_dst);

paddle/operators/load_op.cc

@@ -11,10 +11,10 @@
    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 <fstream>
 
 #include "paddle/framework/op_registry.h"
-
-#include <fstream>
+#include "paddle/platform/device_context.h"
 
 namespace paddle {
 namespace operators {
@@ -26,7 +26,7 @@ class LoadOp : public framework::OperatorBase {
          const framework::AttributeMap &attrs)
       : OperatorBase(type, inputs, outputs, attrs) {}
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &place) const override {
     auto filename = Attr<std::string>("file_path");
     std::ifstream fin(filename);
     PADDLE_ENFORCE(static_cast<bool>(fin), "Cannot open file %s for load op",
@@ -40,7 +40,9 @@ class LoadOp : public framework::OperatorBase {
     auto *tensor = out_var->GetMutable<framework::LoDTensor>();
     framework::DeserializeFromStream(fin, tensor);
 
-    auto place = dev_ctx.GetPlace();
+    platform::DeviceContextPool &pool = platform::DeviceContextPool::Get();
+    auto &dev_ctx = *pool.Borrow(place);
+
     if (platform::is_gpu_place(place)) {
       // copy CPU to GPU
       framework::LoDTensor cpu_tensor;

paddle/operators/lod_array_length_op.cc

@@ -26,7 +26,7 @@ class LoDArrayLengthOp : public framework::OperatorBase {
                    const framework::AttributeMap &attrs)
       : OperatorBase(type, inputs, outputs, attrs) {}
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &place) const override {
     auto &x = scope.FindVar(Input("X"))->Get<framework::LoDTensorArray>();
     auto &out =
         *scope.FindVar(Output("Out"))->GetMutable<framework::LoDTensor>();

paddle/operators/lod_rank_table_op.cc

@@ -24,7 +24,7 @@ class LoDRankTableOp : public framework::OperatorBase {
                  const framework::AttributeMap &attrs)
       : OperatorBase(type, inputs, outputs, attrs) {}
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &dev_place) const override {
     auto x = scope.FindVar(Input("X"))->Get<framework::LoDTensor>();
     auto *out =
         scope.FindVar(Output("Out"))->GetMutable<framework::LoDRankTable>();

paddle/operators/lod_reset_op.cc

@@ -38,7 +38,7 @@ class LoDResetOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::LoDTensor>("X")->type()),
@@ -97,7 +97,7 @@ class LoDResetGradOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::LoDTensor>("X")->type()),

paddle/operators/lod_tensor_to_array_op.cc

@@ -15,6 +15,7 @@
 #include "paddle/framework/lod_tensor_array.h"
 #include "paddle/framework/op_registry.h"
 #include "paddle/operators/detail/safe_ref.h"
+#include "paddle/platform/device_context.h"
 
 namespace paddle {
 namespace operators {
@@ -32,7 +33,7 @@ class LoDTensorToArrayOp : public framework::OperatorBase {
                      const framework::AttributeMap &attrs)
       : OperatorBase(type, inputs, outputs, attrs) {}
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &place) const override {
     auto &x = detail::Ref(scope.FindVar(Input("X")), "Cannot find input %s",
                           Input("X"))
                   .Get<framework::LoDTensor>();
@@ -86,6 +87,10 @@ class LoDTensorToArrayOp : public framework::OperatorBase {
         // out[i][offset: offset+len] = x[each_range.begin: each_range.end]
         auto slice = out[i].Slice(static_cast<int>(offset),
                                   static_cast<int>(offset + len));
+
+        platform::DeviceContextPool &pool = platform::DeviceContextPool::Get();
+        auto &dev_ctx = *pool.Borrow(place);
+
         framework::CopyFrom(x.Slice(static_cast<int>(each_range.begin),
                                     static_cast<int>(each_range.end)),
                             x.place(), dev_ctx, &slice);

paddle/operators/logical_op.cc

@@ -99,9 +99,9 @@ class LogicalOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
-    framework::OpKernelType kt = OperatorWithKernel::GetKernelType(ctx);
+    framework::OpKernelType kt = OperatorWithKernel::GetActualKernelType(ctx);
     // LogicalOp kernel's device type is decided by input tensor place
     kt.place_ = ctx.Input<framework::LoDTensor>("X")->place();
     return kt;

paddle/operators/lookup_table_op.cc

@@ -41,7 +41,7 @@ class LookupTableOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<LoDTensor>("W")->type()),
@@ -98,7 +98,7 @@ class LookupTableOpGrad : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<LoDTensor>("W")->type()),

paddle/operators/lookup_table_op.cu

@@ -101,7 +101,7 @@ class LookupTableGradCUDAKernel : public framework::OpKernel<T> {
       // copy GPU memory to CPU pinned memory
       framework::Vector<int64_t> new_rows;
       new_rows.resize(ids_dim[0]);
-      auto gpu_place = boost::get<platform::GPUPlace>(context.GetPlace());
+      auto gpu_place = boost::get<platform::CUDAPlace>(context.GetPlace());
 
       memory::Copy(platform::CPUPlace(), new_rows.data(), gpu_place, ids_data,
                    ids_dim[0] * sizeof(int64_t), stream);

paddle/operators/lstm_op.cc

@@ -92,7 +92,7 @@ class LSTMOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::LoDTensor>("Input")->type()),
@@ -260,7 +260,7 @@ class LSTMGradOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::LoDTensor>("Input")->type()),

paddle/operators/lstm_unit_op.cu

@@ -98,7 +98,7 @@ class LstmUnitOpCUDAKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
+                   "It must use CUDAPlace.");
 
     auto* x_tensor = ctx.Input<framework::Tensor>("X");
     auto* c_prev_tensor = ctx.Input<framework::Tensor>("C_prev");
@@ -129,7 +129,7 @@ class LstmUnitGradOpCUDAKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
+                   "It must use CUDAPlace.");
 
     auto x_tensor = ctx.Input<Tensor>("X");
     auto c_prev_tensor = ctx.Input<Tensor>("C_prev");

paddle/operators/math/im2col_test.cc

@@ -159,6 +159,7 @@ void testIm2col() {
 TEST(math, im2col) {
   testIm2col<paddle::platform::CPUDeviceContext, paddle::platform::CPUPlace>();
 #ifdef PADDLE_WITH_CUDA
-  testIm2col<paddle::platform::CUDADeviceContext, paddle::platform::GPUPlace>();
+  testIm2col<paddle::platform::CUDADeviceContext,
+             paddle::platform::CUDAPlace>();
 #endif
 }

paddle/operators/math/math_function.cc

@@ -277,14 +277,6 @@ void set_constant_with_place<platform::CPUPlace>(
                            TensorSetConstantCPU(tensor, value));
 }
 
-template <>
-void set_constant_with_place<platform::MKLDNNPlace>(
-    const platform::DeviceContext& context, framework::Tensor* tensor,
-    float value) {
-  framework::VisitDataType(framework::ToDataType(tensor->type()),
-                           TensorSetConstantCPU(tensor, value));
-}
-
 struct TensorSetConstantWithPlace : public boost::static_visitor<void> {
   TensorSetConstantWithPlace(const platform::DeviceContext& context,
                              framework::Tensor* tensor, float value)

paddle/operators/math/math_function.cu

@@ -105,7 +105,7 @@ void matmul<platform::CUDADeviceContext, float>(
   PADDLE_ENFORCE(platform::is_gpu_place(matrix_a.place()) &&
                      platform::is_gpu_place(matrix_b.place()) &&
                      platform::is_gpu_place(matrix_out->place()),
-                 "Matrix must all be in GPUPlace");
+                 "Matrix must all be in CUDAPlace");
 
   int M = dim_out[0];
   int N = dim_out[1];
@@ -134,7 +134,7 @@ void matmul<platform::CUDADeviceContext, double>(
   PADDLE_ENFORCE(platform::is_gpu_place(matrix_a.place()) &&
                      platform::is_gpu_place(matrix_b.place()) &&
                      platform::is_gpu_place(matrix_out->place()),
-                 "Matrix must all be in GPUPlace");
+                 "Matrix must all be in CUDAPlace");
 
   int M = dim_out[0];
   int N = dim_out[1];
@@ -266,20 +266,13 @@ struct TensorSetConstantGPU {
 };
 
 template <>
-void set_constant_with_place<platform::GPUPlace>(
+void set_constant_with_place<platform::CUDAPlace>(
     const platform::DeviceContext& context, framework::Tensor* tensor,
     float value) {
   framework::VisitDataType(framework::ToDataType(tensor->type()),
                            TensorSetConstantGPU(context, tensor, value));
 }
 
-template <>
-void set_constant_with_place<platform::CUDNNPlace>(
-    const platform::DeviceContext& context, framework::Tensor* tensor,
-    float value) {
-  set_constant_with_place<platform::GPUPlace>(context, tensor, value);
-}
-
 template struct RowwiseAdd<platform::CUDADeviceContext, float>;
 template struct RowwiseAdd<platform::CUDADeviceContext, double>;
 template struct ColwiseSum<platform::CUDADeviceContext, float>;

paddle/operators/math/math_function_impl.h

@@ -94,8 +94,8 @@ class ColwiseSum<platform::CPUDeviceContext, T> {
     T* out_buf = out->mutable_data<T>(out->place());
     const T* in_buf = input.data<T>();
 
-    for (size_t i = 0; i < height; ++i) {
-      for (size_t j = 0; j < size; ++j) {
+    for (size_t i = 0; i < static_cast<size_t>(height); ++i) {
+      for (size_t j = 0; j < static_cast<size_t>(size); ++j) {
         if (i == 0) {
           out_buf[j] = in_buf[i * size + j];
         } else {

paddle/operators/math/math_function_test.cu

@@ -13,7 +13,7 @@ TEST(math_function, notrans_mul_trans) {
   float arr[6] = {0, 1, 2, 3, 4, 5};
   memcpy(input1_ptr, arr, 6 * sizeof(float));
 
-  auto* gpu_place = new paddle::platform::GPUPlace(0);
+  auto* gpu_place = new paddle::platform::CUDAPlace(0);
   paddle::platform::CUDADeviceContext context(*gpu_place);
 
   paddle::framework::CopyFrom(input1, *gpu_place, context, &input1_gpu);
@@ -47,7 +47,7 @@ TEST(math_function, trans_mul_notrans) {
   float arr[6] = {0, 1, 2, 3, 4, 5};
   memcpy(input1_ptr, arr, 6 * sizeof(float));
 
-  auto* gpu_place = new paddle::platform::GPUPlace(0);
+  auto* gpu_place = new paddle::platform::CUDAPlace(0);
   paddle::platform::CUDADeviceContext context(*gpu_place);
 
   paddle::framework::CopyFrom(input1, *gpu_place, context, &input1_gpu);
@@ -96,7 +96,7 @@ TEST(math_function, gemm_notrans_cublas) {
   float arr3[8] = {0, 1, 2, 3, 4, 5, 6, 7};
   memcpy(input3_ptr, arr3, 8 * sizeof(float));
 
-  auto* gpu_place = new paddle::platform::GPUPlace(0);
+  auto* gpu_place = new paddle::platform::CUDAPlace(0);
   paddle::platform::CUDADeviceContext context(*gpu_place);
 
   paddle::framework::CopyFrom(input1, *gpu_place, context, &input1_gpu);
@@ -151,7 +151,7 @@ TEST(math_function, gemm_trans_cublas) {
   float arr3[8] = {0, 1, 2, 3, 4, 5, 6, 7};
   memcpy(input3_ptr, arr3, 8 * sizeof(float));
 
-  auto* gpu_place = new paddle::platform::GPUPlace(0);
+  auto* gpu_place = new paddle::platform::CUDAPlace(0);
   paddle::platform::CUDADeviceContext context(*gpu_place);
 
   paddle::framework::CopyFrom(input1, *gpu_place, context, &input1_gpu);
@@ -189,7 +189,7 @@ void GemvTest(int m, int n, bool trans) {
   T* data_b = vec_b.mutable_data<T>({trans ? m : n}, *cpu_place);
   T* data_c = vec_c.mutable_data<T>({trans ? n : m}, *cpu_place);
 
-  auto* gpu_place = new paddle::platform::GPUPlace(0);
+  auto* gpu_place = new paddle::platform::CUDAPlace(0);
   paddle::framework::Tensor g_mat_a;
   paddle::framework::Tensor g_vec_b;
   paddle::framework::Tensor g_vec_c;

paddle/operators/math/selected_rows_functor.cu

@@ -58,15 +58,15 @@ struct SelectedRowsAdd<platform::CUDADeviceContext, T> {
     PADDLE_ENFORCE(platform::is_gpu_place(out_place));
 
     memory::Copy(
-        boost::get<platform::GPUPlace>(out_place), out_data,
-        boost::get<platform::GPUPlace>(in1_place), in1_data,
+        boost::get<platform::CUDAPlace>(out_place), out_data,
+        boost::get<platform::CUDAPlace>(in1_place), in1_data,
         in1_value.numel() * sizeof(T),
         reinterpret_cast<const platform::CUDADeviceContext&>(context).stream());
 
     auto* in2_data = in2_value.data<T>();
-    memory::Copy(boost::get<platform::GPUPlace>(out_place),
+    memory::Copy(boost::get<platform::CUDAPlace>(out_place),
                  out_data + in1_value.numel(),
-                 boost::get<platform::GPUPlace>(in2_place), in2_data,
+                 boost::get<platform::CUDAPlace>(in2_place), in2_data,
                  in2_value.numel() * sizeof(T), context.stream());
   }
 };
@@ -160,9 +160,9 @@ struct SelectedRowsAddTo<platform::CUDADeviceContext, T> {
 
     auto* in1_data = in1_value.data<T>();
     auto* in2_data = in2_value->data<T>();
-    memory::Copy(boost::get<platform::GPUPlace>(in2_place),
+    memory::Copy(boost::get<platform::CUDAPlace>(in2_place),
                  in2_data + input2_offset,
-                 boost::get<platform::GPUPlace>(in1_place), in1_data,
+                 boost::get<platform::CUDAPlace>(in1_place), in1_data,
                  in1_value.numel() * sizeof(T), context.stream());
   }
 };

paddle/operators/math/selected_rows_functor_test.cu

@@ -21,7 +21,7 @@ TEST(selected_rows_functor, gpu_add) {
   using namespace paddle::platform;
   using namespace paddle::operators::math;
 
-  GPUPlace gpu_place(0);
+  CUDAPlace gpu_place(0);
   CPUPlace cpu_place;
   CUDADeviceContext ctx(gpu_place);
   SetConstant<CUDADeviceContext, float> functor;
@@ -119,7 +119,7 @@ TEST(selected_rows_functor, gpu_add_to) {
   using namespace paddle::platform;
   using namespace paddle::operators::math;
 
-  GPUPlace gpu_place(0);
+  CUDAPlace gpu_place(0);
   CPUPlace cpu_place;
   CUDADeviceContext ctx(gpu_place);
   SetConstant<CUDADeviceContext, float> functor;

paddle/operators/math/vol2col_test.cc

@@ -122,6 +122,6 @@ TEST(math, vol2col) {
   testVol2col<paddle::platform::CPUDeviceContext, paddle::platform::CPUPlace>();
 #ifdef PADDLE_WITH_CUDA
   testVol2col<paddle::platform::CUDADeviceContext,
-              paddle::platform::GPUPlace>();
+              paddle::platform::CUDAPlace>();
 #endif  // PADDLE_WITH_CUDA
 }

paddle/operators/max_sequence_len_op.cc

@@ -28,7 +28,7 @@ class MaxSeqenceLenOp : public framework::OperatorBase {
       : OperatorBase(type, inputs, outputs, attrs) {}
 
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &dev_place) const override {
     auto &rank_table =
         scope.FindVar(Input("RankTable"))->Get<framework::LoDRankTable>();
     auto *out =

paddle/operators/merge_lod_tensor_op.cc

@@ -28,7 +28,11 @@ class MergeLoDTensorOp : public framework::OperatorBase {
                    const framework::AttributeMap &attrs)
       : OperatorBase(type, inputs, outputs, attrs) {}
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &dev_place) const override {
+    // get device context from pool
+    platform::DeviceContextPool &pool = platform::DeviceContextPool::Get();
+    auto &dev_ctx = *pool.Borrow(dev_place);
+
     auto &x = scope.FindVar(Input("X"))->Get<framework::LoDTensor>();
     auto &mask = scope.FindVar(Input("Mask"))->Get<framework::LoDTensor>();
     auto &in_true = scope.FindVar(Input("InTrue"))->Get<framework::LoDTensor>();

paddle/operators/mul_op.cc

@@ -113,7 +113,7 @@ This operator is used to perform matrix multiplication for input $X$ and $Y$.
 
 The equation is:
 
-    $$Out = X * Y$$
+$$Out = X * Y$$
 
 Both the input $X$ and $Y$ can carry the LoD (Level of Details) information,
 or not. But the output only shares the LoD information with input $X$.

paddle/operators/multiplex_op.cc

@@ -51,7 +51,7 @@ class MultiplexOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.MultiInput<Tensor>("X")[0]->type()),
@@ -102,7 +102,7 @@ class MultiplexGradOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.MultiInput<Tensor>("X")[0]->type()),

paddle/operators/multiplex_op.cu

@@ -36,7 +36,7 @@ class MultiplexGPUKernel : public framework::OpKernel<T> {
     CopyFrom(*ids, platform::CPUPlace(), ctx.device_context(), &index_t_cpu);
     auto* index = index_t_cpu.data<int32_t>();
     auto stream = ctx.cuda_device_context().stream();
-    platform::GPUPlace place = boost::get<platform::GPUPlace>(ctx.GetPlace());
+    platform::CUDAPlace place = boost::get<platform::CUDAPlace>(ctx.GetPlace());
     for (auto i = 0; i < rows; i++) {
       int32_t k = index[i];
       PADDLE_ENFORCE_GE(k, 0, "index must be nonnegative.");
@@ -73,7 +73,7 @@ class MultiplexGradGPUKernel : public framework::OpKernel<T> {
     auto* index = index_t_cpu.data<int32_t>();
 
     auto stream = ctx.cuda_device_context().stream();
-    platform::GPUPlace place = boost::get<platform::GPUPlace>(ctx.GetPlace());
+    platform::CUDAPlace place = boost::get<platform::CUDAPlace>(ctx.GetPlace());
     for (auto i = 0; i < rows; i++) {
       size_t k = static_cast<size_t>(index[i]);
       if (d_ins[k]) {

paddle/operators/nccl_op.cc

@@ -24,7 +24,7 @@ class NCCLInitOp : public framework::OperatorBase {
       : OperatorBase(type, inputs, outputs, attrs) {}
 
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &place) const override {
     const auto &name = Output("Communicator");
     PADDLE_ENFORCE_NOT_NULL(scope.FindVar(name),
                             "Can not find variable '%s' in the scope.", name);

paddle/operators/nccl_op.cu.cc

@@ -67,7 +67,7 @@ class NCCLAllReduceKernel : public framework::OpKernel<T> {
     auto stream = ctx.cuda_device_context().stream();
 
     // device id
-    int gpu_id = boost::get<platform::GPUPlace>(ctx.GetPlace()).GetDeviceId();
+    int gpu_id = boost::get<platform::CUDAPlace>(ctx.GetPlace()).GetDeviceId();
     int idx = comm->GetCommId(gpu_id);
 
     for (size_t i = 0; i < ins.size(); ++i) {
@@ -120,7 +120,7 @@ class NCCLReduceKernel : public framework::OpKernel<T> {
                       ctx.device_context())
                       .stream();
     // device id
-    int gpu_id = boost::get<platform::GPUPlace>(ctx.GetPlace()).GetDeviceId();
+    int gpu_id = boost::get<platform::CUDAPlace>(ctx.GetPlace()).GetDeviceId();
     int idx = comm->GetCommId(gpu_id);
 
     auto ins_names = ctx.Inputs("X");
@@ -164,7 +164,7 @@ class NCCLBcastKernel : public framework::OpKernel<T> {
                       ctx.device_context())
                       .stream();
     // device id
-    int gpu_id = boost::get<platform::GPUPlace>(ctx.GetPlace()).GetDeviceId();
+    int gpu_id = boost::get<platform::CUDAPlace>(ctx.GetPlace()).GetDeviceId();
     int idx = comm->GetCommId(gpu_id);
 
     if (idx == root) {

paddle/operators/nccl_op_test.cu.cc

@@ -22,6 +22,7 @@
 #include <vector>
 
 #include "paddle/framework/block_desc.h"
+#include "paddle/framework/init.h"
 #include "paddle/framework/op_desc.h"
 #include "paddle/framework/op_registry.h"
 #include "paddle/framework/program_desc.h"
@@ -49,9 +50,9 @@ const f::DDim kDims = {100, 100};
 class NCCLTester : public ::testing::Test {
  public:
   virtual void SetUp() override {
-    cpu_ctx = new p::CPUDeviceContext(p::CPUPlace());
+    paddle::platform::CPUPlace cpu_place;
     for (size_t i = 0; i < gpu_list.size(); ++i) {
-      p::GPUPlace place(i);
+      p::CUDAPlace place(i);
       dev_ctxs.emplace_back(new p::CUDADeviceContext(place));
     }
 
@@ -65,6 +66,7 @@ class NCCLTester : public ::testing::Test {
   }
 
   void NCCLInitOp() {
+    paddle::platform::CPUPlace cpu_place;
     std::unique_ptr<f::OpDesc> op1(new f::OpDesc);
 
     op1->SetType("ncclInit");
@@ -76,7 +78,7 @@ class NCCLTester : public ::testing::Test {
 
     auto op = f::OpRegistry::CreateOp(*op1);
     VLOG(1) << "invoke NCCLInitOp.";
-    op->Run(g_scope, *cpu_ctx);
+    op->Run(g_scope, cpu_place);
     VLOG(1) << "NCCLInitOp finished.";
   }
 
@@ -85,7 +87,7 @@ class NCCLTester : public ::testing::Test {
     std::unique_lock<std::mutex> lk(mu);
     const f::OpDesc *op1 = &op_desc;
 
-    p::GPUPlace place(gpu_id);
+    p::CUDAPlace place(gpu_id);
     auto &ctx = dev_ctxs.at(gpu_id);
 
     auto *send_tensor = scope->Var("st")->GetMutable<f::LoDTensor>();
@@ -111,13 +113,12 @@ class NCCLTester : public ::testing::Test {
     VLOG(1) << "Device : " << gpu_id << " invoke " << op_desc.Type();
     VLOG(1) << " send_tensor : " << send_tensor->numel()
             << " recv_tensor : " << recv_tensor->numel();
-    op->Run(*scope, *ctx);
+    op->Run(*scope, place);
     VLOG(1) << "Device : " << gpu_id << " finished " << op_desc.Type();
   }
 
  public:
   std::vector<p::DeviceContext *> dev_ctxs;
-  p::DeviceContext *cpu_ctx;
   f::Scope g_scope;
   std::mutex mu;
 };
@@ -131,14 +132,14 @@ TEST(NCCL, ncclInitOp) {
   op_desc->SetAttr("gpus", {gpu_list});
 
   f::Scope g_scope;
-  std::unique_ptr<p::DeviceContext> ctx(new p::CPUDeviceContext(p::CPUPlace()));
+  paddle::platform::CPUPlace cpu_place;
 
   auto *var = g_scope.Var("x1");
   var->GetMutable<p::Communicator>();
 
   auto op = f::OpRegistry::CreateOp(*op_desc);
   VLOG(1) << "invoke NCCLInitOp.";
-  op->Run(g_scope, *ctx.get());
+  op->Run(g_scope, cpu_place);
   VLOG(1) << "NCCLInitOp finished.";
 }
 
@@ -170,7 +171,7 @@ TEST_F(NCCLTester, ncclAllReduceOp) {
 
   for (size_t i = 0; i < dev_scopes.size(); ++i) {
     p::CPUPlace cpu_place;
-    p::GPUPlace gpu_place(gpu_list[i]);
+    p::CUDAPlace gpu_place(gpu_list[i]);
 
     auto &recv_tensor = dev_scopes[i]->FindVar("rt")->Get<f::LoDTensor>();
     auto *rt = recv_tensor.data<float>();
@@ -179,7 +180,7 @@ TEST_F(NCCLTester, ncclAllReduceOp) {
     auto *ct = result_tensor->mutable_data<float>(cpu_place);
 
     paddle::memory::Copy(
-        cpu_place, ct, p::GPUPlace(gpu_list[i]), rt,
+        cpu_place, ct, p::CUDAPlace(gpu_list[i]), rt,
         recv_tensor.numel() * sizeof(float),
         static_cast<p::CUDADeviceContext *>(dev_ctxs[i])->stream());
 
@@ -218,7 +219,7 @@ TEST_F(NCCLTester, ncclReduceOp) {
   float result = std::accumulate(gpu_list.begin(), gpu_list.end(), 0);
 
   p::CPUPlace cpu_place;
-  p::GPUPlace gpu_place(gpu_list[kRoot]);
+  p::CUDAPlace gpu_place(gpu_list[kRoot]);
 
   auto &recv_tensor = dev_scopes[kRoot]->FindVar("rt")->Get<f::LoDTensor>();
   auto *rt = recv_tensor.data<float>();
@@ -228,7 +229,7 @@ TEST_F(NCCLTester, ncclReduceOp) {
   auto *ct = result_tensor->mutable_data<float>(cpu_place);
 
   paddle::memory::Copy(
-      cpu_place, ct, p::GPUPlace(gpu_list[kRoot]), rt,
+      cpu_place, ct, p::CUDAPlace(gpu_list[kRoot]), rt,
       recv_tensor.numel() * sizeof(float),
       static_cast<p::CUDADeviceContext *>(dev_ctxs[kRoot])->stream());
 
@@ -267,7 +268,7 @@ TEST_F(NCCLTester, ncclBcastOp) {
   float result = kRoot;
 
   p::CPUPlace cpu_place;
-  p::GPUPlace gpu_place(gpu_list[idx]);
+  p::CUDAPlace gpu_place(gpu_list[idx]);
 
   auto &recv_tensor = dev_scopes[idx]->FindVar("rt")->Get<f::LoDTensor>();
   auto *rt = recv_tensor.data<float>();
@@ -276,7 +277,7 @@ TEST_F(NCCLTester, ncclBcastOp) {
   auto *ct = result_tensor->mutable_data<float>(cpu_place);
 
   paddle::memory::Copy(
-      cpu_place, ct, p::GPUPlace(gpu_list[idx]), rt,
+      cpu_place, ct, p::CUDAPlace(gpu_list[idx]), rt,
       recv_tensor.numel() * sizeof(float),
       static_cast<p::CUDADeviceContext *>(dev_ctxs[idx])->stream());
 
@@ -294,9 +295,18 @@ int main(int argc, char **argv) {
     return 0;
   }
 
-  for (int i = 0; i < dev_count; ++i) {
+  std::vector<paddle::platform::Place> places;
+
+  places.emplace_back(paddle::platform::CPUPlace());
+  int count = paddle::platform::GetCUDADeviceCount();
+  for (int i = 0; i < count; ++i) {
+    places.emplace_back(paddle::platform::CUDAPlace(i));
     gpu_list.emplace_back(i);
   }
+
+  VLOG(0) << " DeviceCount " << count;
+  paddle::platform::DeviceContextPool::Create(places);
+
   testing::InitGoogleTest(&argc, argv);
 
   // device context should be release before scope.

paddle/operators/nce_op.cc

@@ -63,7 +63,7 @@ class NCEOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("Input")->type()),
@@ -166,7 +166,7 @@ class NCEOpGrad : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("Input")->type()),

paddle/operators/net_op.h

@@ -65,9 +65,9 @@ class NetOp : public framework::OperatorBase {
    * will be used.
    */
   void Run(const framework::Scope& scope,
-           const platform::DeviceContext& dev_ctx) const override {
+           const platform::Place& place) const override {
     for (auto& op : ops_) {
-      op->Run(scope, dev_ctx);
+      op->Run(scope, place);
     }
   }
 

paddle/operators/net_op_test.cc

@@ -13,8 +13,7 @@ class TestOp : public framework::OperatorBase {
  public:
   using framework::OperatorBase::OperatorBase;
   DEFINE_OP_CLONE_METHOD(TestOp);
-  void Run(const Scope& scope,
-           const platform::DeviceContext& dev_ctx) const override {
+  void Run(const Scope& scope, const platform::Place& place) const override {
     ++run_cnt;
   }
 };

paddle/operators/pool_cudnn_op.cu.cc

@@ -29,7 +29,7 @@ class PoolCudnnOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
+                   "It must use CUDAPlace.");
 
     const Tensor *input = ctx.Input<Tensor>("X");
     Tensor *output = ctx.Output<Tensor>("Out");
@@ -90,7 +90,7 @@ class PoolCudnnGradOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
+                   "It must use CUDAPlace.");
 
     const Tensor *input = ctx.Input<Tensor>("X");
     const Tensor *output = ctx.Input<Tensor>("Out");

paddle/operators/pool_with_index_op.cc

@@ -69,7 +69,7 @@ class MaxPoolWithIndexOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),
@@ -90,7 +90,7 @@ class MaxPoolWithIndexOpGrad : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),

paddle/operators/positive_negative_pair_op.cc

@@ -85,7 +85,7 @@ class PositiveNegativePairOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("Score")->type()),
@@ -154,13 +154,14 @@ class PositiveNegativePairOpMaker : public framework::OpProtoAndCheckerMaker {
         "Noting that reducing on the first dim will make the LoD info lost.")
         .SetDefault(0);
     AddComment(R"DOC(
-        PositiveNegativePairOp can be used to evaluate Learning To Rank(LTR) 
-        model performance. 
-        Within some context, e.g. the "query", a LTR model generates scores
-        for a list of items, which gives a partial order of the items.
-        PositiveNegativePairOp takes a list of reference rank order 
-        (Input("Label")) and the model generated scores (Input(Score)) as 
-        inputs and counts the pairs that ranked correctly and incorrectly.
+PositiveNegativePairOp can be used to evaluate Learning To Rank(LTR) model's
+performance.
+
+Within some context, e.g. the "query", a LTR model generates scores for a list
+of items, which gives a partial order of the items. PositiveNegativePairOp
+takes a list of reference rank order (Input("Label")) and the model generated
+scores (Input(Score)) as inputs and counts the pairs that ranked correctly
+and incorrectly.
 )DOC");
   }
 };

paddle/operators/precision_recall_op.cc

@@ -80,7 +80,7 @@ class PrecisionRecallOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("MaxProbs")->type()),

paddle/operators/recurrent_op.cc

@@ -227,14 +227,15 @@ class RecurrentOp : public RecurrentBase {
       : RecurrentBase(type, inputs, outputs, attrs) {}
 
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &place) const override {
     auto seq_len = static_cast<size_t>(this->GetSequenceLength(scope));
     VLOG(3) << "Static RNN input sequence length = " << seq_len;
     StepScopes scopes = CreateStepScopes(scope, seq_len);
     auto reverse = Attr<bool>(kReverse);
 
-    framework::Executor executor(dev_ctx);
+    framework::Executor executor(place);
     auto *block = Attr<framework::BlockDesc *>(kStepBlock);
+
     auto *program = block->Program();
 
     for (size_t i = 0; i < seq_len; ++i) {
@@ -270,6 +271,10 @@ class RecurrentOp : public RecurrentBase {
       executor.Run(*program, &cur_scope, block->ID(),
                    false /*create_local_scope*/);
 
+      // get device context from pool
+      platform::DeviceContextPool &pool = platform::DeviceContextPool::Get();
+      auto &dev_ctx = *pool.Borrow(place);
+
       // Copy inside::output -> outside::output
       //    outside::output[seq_offset: seq_offset + 1] = inside::output
       this->LinkTensorWithCallback(
@@ -278,14 +283,13 @@ class RecurrentOp : public RecurrentBase {
               framework::LoDTensor *dst_tensor) {
             if (i == 0) {  // create output tensor at begin
               dst_tensor->Resize(PrependDims(seq_len, src_tensor.dims()));
-              dst_tensor->mutable_data(dev_ctx.GetPlace(), src_tensor.type());
+              dst_tensor->mutable_data(place, src_tensor.type());
             }
 
             auto dst_out = dst_tensor->Slice(seq_offset, seq_offset + 1);
             // Explicit copy output since the local RNN scope can be destroyed
             // early.
-            framework::CopyFrom(src_tensor, dev_ctx.GetPlace(), dev_ctx,
-                                &dst_out);
+            framework::CopyFrom(src_tensor, place, dev_ctx, &dst_out);
           });
 
       scopes.Next();
@@ -311,15 +315,20 @@ class RecurrentGradOp : public RecurrentBase {
       : RecurrentBase(type, inputs, outputs, attrs) {}
 
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &place) const override {
     auto seq_len = static_cast<size_t>(GetSequenceLength(scope));
     StepScopes scopes = CreateStepScopes(scope, seq_len);
     auto reverse = Attr<bool>(kReverse);
 
-    framework::Executor executor(dev_ctx);
+    framework::Executor executor(place);
     auto *block = Attr<framework::BlockDesc *>(kStepBlock);
+
     auto *program = block->Program();
 
+    // get device context from pool
+    platform::DeviceContextPool &pool = platform::DeviceContextPool::Get();
+    auto &dev_ctx = *pool.Borrow(place);
+
     for (size_t step_id = 0; step_id < seq_len; ++step_id) {
       size_t seq_offset = reverse ? step_id : seq_len - step_id - 1;
       VLOG(3) << "Recurrent backward operate at the time step " << seq_offset;
@@ -366,8 +375,7 @@ class RecurrentGradOp : public RecurrentBase {
           auto *cur_grad_var = cur_scope.Var(cur_grad);
           auto cur_grad_tensor =
               cur_grad_var->GetMutable<framework::LoDTensor>();
-          framework::CopyFrom(ex_tensor, dev_ctx.GetPlace(), dev_ctx,
-                              cur_grad_tensor);
+          framework::CopyFrom(ex_tensor, place, dev_ctx, cur_grad_tensor);
         }
       }
 
@@ -410,7 +418,7 @@ class RecurrentGradOp : public RecurrentBase {
             auto zero_op = framework::OpRegistry::CreateOp(
                 "fill_constant", framework::VariableNameMap{},
                 {{"Out", {pg_names[param_id]}}}, attrs);
-            zero_op->Run(scope, dev_ctx);
+            zero_op->Run(scope, place);
           }
 
           auto new_inside_name = cur_scope.Rename(inside_grad_name);
@@ -419,7 +427,7 @@ class RecurrentGradOp : public RecurrentBase {
           auto sum_op = framework::OpRegistry::CreateOp(
               "sum", {{"X", {pg_names[param_id], new_inside_name}}},
               {{"Out", {pg_names[param_id]}}}, framework::AttributeMap{});
-          sum_op->Run(cur_scope, dev_ctx);
+          sum_op->Run(cur_scope, place);
 
           cur_scope.Rename(new_inside_name, inside_grad_name);
         }
@@ -437,11 +445,11 @@ class RecurrentGradOp : public RecurrentBase {
             }
             if (step_id == 0) {  // alloc memory
               outside->Resize(PrependDims(seq_len, inside.dims()));
-              outside->mutable_data(dev_ctx.GetPlace(), inside.type());
+              outside->mutable_data(place, inside.type());
             }
 
             auto dst = outside->Slice(seq_offset, seq_offset + 1);
-            framework::CopyFrom(inside, dev_ctx.GetPlace(), dev_ctx, &dst);
+            framework::CopyFrom(inside, place, dev_ctx, &dst);
           });
       VLOG(5) << "Link outside gradient finished ";
 
@@ -453,8 +461,8 @@ class RecurrentGradOp : public RecurrentBase {
             [&](const framework::LoDTensor &inside,
                 framework::LoDTensor *outside) {
               outside->Resize(inside.dims());
-              outside->mutable_data(dev_ctx.GetPlace(), inside.type());
-              framework::CopyFrom(inside, dev_ctx.GetPlace(), dev_ctx, outside);
+              outside->mutable_data(place, inside.type());
+              framework::CopyFrom(inside, place, dev_ctx, outside);
             });
         VLOG(5) << "Link initialize state gradient finished ";
       }
@@ -570,7 +578,7 @@ class RecurrentGradOpDescMaker : public framework::SingleGradOpDescMaker {
     for (auto &input_param : this->InputNames()) {
       grad->SetInput(input_param, this->Input(input_param));
       grad->SetOutput(framework::GradVarName(input_param),
-                      this->InputGrad(input_param));
+                      this->InputGrad(input_param, false));
     }
 
     for (auto &output_param : this->OutputNames()) {

paddle/operators/recv_op.cc

@@ -24,6 +24,7 @@
 #include "paddle/framework/framework.pb.h"
 #include "paddle/framework/lod_tensor.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/framework/proto_desc.h"
 #include "paddle/operators/detail/send_recv_impl.h"
 #include "paddle/operators/detail/simple_block_queue.h"
 
@@ -61,29 +62,78 @@ class RecvOp : public framework::OperatorBase {
     server_thread_->join();
   }
 
+  std::string GetGradVarNameForTrainer(const std::string &varname) const {
+    if (grads_counter_.find(varname) == grads_counter_.end()) {
+      grads_counter_[varname] = 0;
+    }
+    char ret[256];
+    snprintf(ret, sizeof(ret), "%s.trainer_%d", varname.c_str(),
+             grads_counter_[varname]++);
+    return std::string(ret);
+  }
+
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
-    // blocking get one var from client.
-    const framework::LoDTensor &t = rpc_service_->Get();
+           const platform::Place &dev_place) const override {
+    // FIXME(typhoonzero): no new scopes for every run.
     framework::Scope &recv_scope = scope.NewScope();
-    // set graph input var
-    auto *var = recv_scope.Var(Input("RX"));
+    rpc_service_->SetScope(&recv_scope);
+    auto param_list = Attr<std::vector<std::string>>("ParamList");
+    auto grad_list = Attr<std::vector<std::string>>("GradList");
+    auto trainer_count = Attr<int>("Trainers");
+    size_t param_count = param_list.size();
+    rpc_service_->Reset();
+    // TODO(typhoonzero): change this to a while_op for every cluster-batch.
+    while (true) {
+      // Get from multiple trainers, we don't care about order in which
+      // the gradient arrives, just add suffix 0~n then average the gradient.
+      for (size_t i = 0; i < param_count * trainer_count; ++i) {
+        // blocking get one var from client.
+        const detail::TensorWithName &v = rpc_service_->Get();
+        auto grad_var_name = v.first;
+        auto it = std::find(grad_list.begin(), grad_list.end(), grad_var_name);
+        std::string param_var_name;
+        if (it != grad_list.end()) {
+          param_var_name = param_list[it - grad_list.begin()];
+        } else {
+          LOG(ERROR) << "grad have no paired param found!";
+        }
+        VLOG(3) << "recved grad: " << grad_var_name
+                << " updating param: " << param_var_name;
+        auto *merged_grad = recv_scope.FindVar(grad_var_name);
+        if (merged_grad == nullptr) {
+          // create output of merged var.
+          auto merged_var = recv_scope.Var(grad_var_name);
+          merged_var->GetMutable<framework::LoDTensor>();
+        }
+
+        if (trainer_count > 1) {
+          grad_var_name = this->GetGradVarNameForTrainer(grad_var_name);
+        }
+
+        auto *var = recv_scope.Var(grad_var_name);
         auto *tensor = var->GetMutable<framework::LoDTensor>();
         // FIXME(typhoonzero): do not copy
-    framework::CopyFrom(t, dev_ctx.GetPlace(), dev_ctx, tensor);
+        platform::DeviceContextPool &pool = platform::DeviceContextPool::Get();
+        auto &dev_ctx = *pool.Borrow(place);
+        framework::CopyFrom(v.second, place, dev_ctx, tensor);
+      }
+      rpc_service_->Reset();
 
       std::string program_str = Attr<std::string>("OptimizeProgram");
-    framework::ProgramDesc program_desc;
+      framework::proto::ProgramDesc program_desc;
       program_desc.ParseFromString(program_str);
-    framework::ProgramDescBind program(program_desc);
-    framework::Executor executor(dev_ctx);
+      framework::ProgramDesc program(program_desc);
+      framework::Executor executor(place);
       // Run sub graph to get optimized tensor
+      try {
         executor.Run(program, &recv_scope, 0, /*global_block*/
-                 false /*create_local_scope*/);
-
-    auto *out_var = recv_scope.FindVar("Out");
-    // push back
-    rpc_service_->Push(out_var->Get<framework::LoDTensor>());
+                     false /*create_local_scope*/, false /*create_vars*/);
+      } catch (std::exception &e) {
+        LOG(ERROR) << "run sub program error " << e.what();
+      }
+      rpc_service_->Done();
+      grads_counter_.clear();
+    }  // while(true)
   }
 
  protected:
@@ -93,13 +143,14 @@ class RecvOp : public framework::OperatorBase {
   // grpc send/recv service implement to register.
   std::shared_ptr<detail::SendRecvServerImpl> rpc_service_;
   std::shared_ptr<std::thread> server_thread_;
+  mutable std::unordered_map<std::string, int> grads_counter_;
 };
 
 class RecvOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   RecvOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("RX", "(Tensor) Input tensor to be saved");
+    AddInput("RX", "(Tensor) Input tensor to be optimized").AsDuplicable();
     AddComment(R"DOC(
 Recv operator
 
@@ -112,6 +163,17 @@ This operator will recv tensor from send_op
         .AddCustomChecker([](const std::string &ip) { return !ip.empty(); });
     AddAttr<std::string>("OptimizeProgram", "type string",
                          "Serialized ProgramDesc string for recv to run.");
+    AddAttr<std::vector<std::string>>(
+        "ParamList", "type list of string",
+        "grad->param name mapping to find which param to optimize.")
+        .SetDefault({});
+    AddAttr<std::vector<std::string>>(
+        "GradList", "type list of string",
+        "grad->param name mapping to find which param to optimize.")
+        .SetDefault({});
+    AddAttr<int>("Trainers", "type int",
+                 "Number of trainers in the current cluster job")
+        .SetDefault(1);
   }
 };
 

paddle/operators/reorder_lod_tensor_by_rank_op.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/lod_rank_table.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/detail/safe_ref.h"
+#include "paddle/platform/device_context.h"
+
+namespace paddle {
+namespace operators {
+
+class ReorderLoDTensorByRankTableOpProtoMaker
+    : public framework::OpProtoAndCheckerMaker {
+ public:
+  ReorderLoDTensorByRankTableOpProtoMaker(OpProto *proto,
+                                          OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "(LoDTensor) the input lod tensor need to be reordered.");
+    AddInput("RankTable",
+             "(LoDRankTable) the rank table that input need follow");
+    AddOutput("Out", "(LoDTensor) reordered lod tensor");
+    AddComment(R"DOC(ReorderLoDTensorByRankTable
+
+Reorder the input X by the rank of `RankTable`. If `RankTable` is ordered by
+index [3, 0, 2, 1]. Input X will reorder its sequence, the third sequence of
+X will be the first sequence of Output.
+
+NOTE: The RankTable does not need to be calculated by X.
+
+For example:
+The X = [Seq0, Seq1, Seq2, Seq3]. The indices of RankTable are [3, 0, 2, 1].
+
+The Out =  [Seq3, Seq0, Seq2, Seq1] with correct LoD information.
+)DOC");
+  }
+};
+
+class ReorderLoDTensorByRankTableBase : public framework::OperatorBase {
+ public:
+  ReorderLoDTensorByRankTableBase(const std::string &type,
+                                  const framework::VariableNameMap &inputs,
+                                  const framework::VariableNameMap &outputs,
+                                  const framework::AttributeMap &attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {}
+  void Run(const framework::Scope &scope,
+           const platform::Place &place) const override {
+    auto &x =
+        detail::Ref(scope.FindVar(Input("X")),
+                    "Cannot find input lod tensor variable %s", Input("X"))
+            .Get<framework::LoDTensor>();
+    auto &rank_table = detail::Ref(scope.FindVar(Input("RankTable")),
+                                   "Cannot find input rank table variable %s",
+                                   Input("RankTable"))
+                           .Get<framework::LoDRankTable>();
+    auto &out =
+        *detail::Ref(scope.FindVar(Output("Out")),
+                     "Cannot find output lod tensor variable %s", Output("Out"))
+             .GetMutable<framework::LoDTensor>();
+
+    out.Resize(x.dims());
+    out.mutable_data(x.place(), x.type());
+    this->process(place, x, rank_table, &out);
+  }
+
+ protected:
+  virtual void process(const platform::Place &place,
+                       const framework::LoDTensor &x,
+                       const framework::LoDRankTable &rank_table,
+                       framework::LoDTensor *out) const = 0;
+
+  struct AbsoluteRankTableItem {
+    size_t offset;  // the absolute/accumulated offset.
+    size_t length;  // the length
+    framework::LoD lod;
+  };
+
+  std::vector<AbsoluteRankTableItem> GetAbsoluteOffsetAndLengthByLoDRankTable(
+      const framework::LoDTensor &x) const {
+    std::vector<AbsoluteRankTableItem> absolute_table;
+    size_t level = 0;
+    size_t size = x.lod()[level].size();
+
+    for (size_t i = 0; i < size - 1; ++i) {
+      auto lod_offset =
+          framework::GetSubLoDAndAbsoluteOffset(x.lod(), i, i + 1, level);
+
+      auto &offset = lod_offset.second;
+
+      absolute_table.emplace_back();
+      absolute_table.back().length = offset.second - offset.first;
+      absolute_table.back().offset = offset.first;
+      absolute_table.back().lod = lod_offset.first;
+    }
+    return absolute_table;
+  }
+
+  size_t CopyTensorAndLod(const platform::Place &place,
+                          const AbsoluteRankTableItem &item,
+                          const framework::LoDTensor &x,
+                          framework::LoDTensor *out, size_t out_offset) const {
+    auto &out_lod = *out->mutable_lod();
+    auto len = item.length;
+    auto x_offset = item.offset;
+
+    if (out_lod.empty()) {
+      for (size_t i = 0; i < item.lod.size(); ++i) {
+        out_lod.push_back(std::vector<size_t>({0}));
+      }
+    }
+
+    for (size_t i = 0; i < out_lod.size(); ++i) {
+      auto &out_v = out_lod[i];
+      auto &new_lod_v = item.lod[i];
+
+      for (auto &detail : new_lod_v) {
+        out_v.push_back(out_v.back() + detail);
+      }
+    }
+
+    auto x_sliced = x.Slice(x_offset, x_offset + len);
+    auto out_sliced = out->Slice(out_offset, out_offset + len);
+
+    platform::DeviceContextPool &pool = platform::DeviceContextPool::Get();
+    auto &dev_ctx = *pool.Borrow(place);
+    framework::CopyFrom(x_sliced, out_sliced.place(), dev_ctx, &out_sliced);
+    out_offset += len;
+    return out_offset;
+  }
+};
+
+class ReorderLoDTensorByRankTableOp : public ReorderLoDTensorByRankTableBase {
+ public:
+  ReorderLoDTensorByRankTableOp(const std::string &type,
+                                const framework::VariableNameMap &inputs,
+                                const framework::VariableNameMap &outputs,
+                                const framework::AttributeMap &attrs)
+      : ReorderLoDTensorByRankTableBase(type, inputs, outputs, attrs) {}
+
+ protected:
+  void process(const platform::Place &place, const framework::LoDTensor &x,
+               const framework::LoDRankTable &rank_table,
+               framework::LoDTensor *out) const override {
+    auto absolute_table = GetAbsoluteOffsetAndLengthByLoDRankTable(x);
+    size_t out_offset = 0;
+    out->mutable_lod()->clear();
+    for (auto &item : rank_table.items()) {
+      PADDLE_ENFORCE_LT(item.index, absolute_table.size());
+      out_offset = CopyTensorAndLod(place, absolute_table[item.index], x, out,
+                                    out_offset);
+    }
+  }
+};
+
+class IdentityInferShape : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext *context) const override {
+    context->SetOutputDim("Out", context->GetInputDim("X"));
+  }
+};
+
+class ReorderLodTensorByRankGradOpMaker
+    : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDesc> Apply() const override {
+    auto *grad_op = new framework::OpDesc();
+    grad_op->SetType("reorder_lod_tensor_by_rank_grad");
+    grad_op->SetInput("X", OutputGrad("Out"));
+    grad_op->SetOutput("Out", InputGrad("X"));
+    grad_op->SetInput("RankTable", Input("RankTable"));
+    return std::unique_ptr<framework::OpDesc>(grad_op);
+  }
+};
+
+class ReorderLoDTensorByRankGradOp : public ReorderLoDTensorByRankTableBase {
+ public:
+  ReorderLoDTensorByRankGradOp(const std::string &type,
+                               const framework::VariableNameMap &inputs,
+                               const framework::VariableNameMap &outputs,
+                               const framework::AttributeMap &attrs)
+      : ReorderLoDTensorByRankTableBase(type, inputs, outputs, attrs) {}
+
+ protected:
+  void process(const platform::Place &place, const framework::LoDTensor &x,
+               const framework::LoDRankTable &rank_table,
+               framework::LoDTensor *out) const override {
+    auto absolute_table = GetAbsoluteOffsetAndLengthByLoDRankTable(x);
+
+    // offsets = enumerate([item.index for item in rank_table.items()])
+    std::vector<std::pair<size_t, size_t>> offsets;
+    offsets.reserve(rank_table.items().size());
+    for (size_t i = 0; i < rank_table.items().size(); ++i) {
+      offsets.push_back({i, rank_table.items()[i].index});
+    }
+
+    // offsets.sort(key=lambda x: x[1])
+    std::sort(
+        offsets.begin(), offsets.end(),
+        [](const std::pair<size_t, size_t> &a,
+           const std::pair<size_t, size_t> &b) { return a.second < b.second; });
+
+    // Copy TensorAndLod
+    size_t out_offset = 0;
+    for (auto &offset : offsets) {
+      out_offset = this->CopyTensorAndLod(place, absolute_table[offset.first],
+                                          x, out, out_offset);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OPERATOR(reorder_lod_tensor_by_rank,
+                  ops::ReorderLoDTensorByRankTableOp,
+                  ops::ReorderLodTensorByRankGradOpMaker,
+                  ops::ReorderLoDTensorByRankTableOpProtoMaker,
+                  ops::IdentityInferShape);
+REGISTER_OPERATOR(reorder_lod_tensor_by_rank_grad,
+                  ops::ReorderLoDTensorByRankGradOp, ops::IdentityInferShape);

paddle/operators/reshape_op.cu

@@ -16,7 +16,7 @@
 
 REGISTER_OP_CUDA_KERNEL(
     reshape,
-    paddle::operators::ReshapeKernel<paddle::platform::GPUPlace, float>);
+    paddle::operators::ReshapeKernel<paddle::platform::CUDAPlace, float>);
 REGISTER_OP_CUDA_KERNEL(
     reshape_grad,
-    paddle::operators::ReshapeGradKernel<paddle::platform::GPUPlace, float>);
+    paddle::operators::ReshapeGradKernel<paddle::platform::CUDAPlace, float>);

paddle/operators/rnn_memory_helper_op.cc

@@ -25,7 +25,7 @@ class RNNMemoryHelperOp : public framework::OperatorBase {
                     const framework::AttributeMap &attrs)
       : OperatorBase(type, inputs, outputs, attrs) {}
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &dev_place) const override {
     auto mem_var_name = Input("X");
     auto *mem_var = scope.FindVar(mem_var_name);
     PADDLE_ENFORCE(mem_var != nullptr,
@@ -77,7 +77,7 @@ class RNNMemoryHelperGradOp : public framework::OperatorBase {
                         const framework::AttributeMap &attrs)
       : OperatorBase(type, inputs, outputs, attrs) {}
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &dev_place) const override {
     auto out_grad_var_name = Input(framework::GradVarName("Out"));
     auto *out_grad_var = scope.FindVar(out_grad_var_name);
 
@@ -100,7 +100,7 @@ class RNNMemoryHelperGradOp : public framework::OperatorBase {
 
       auto zero_op = framework::OpRegistry::CreateOp(
           "fill_constant", {}, {{"Out", {in_grad_var_name}}}, attrs);
-      zero_op->Run(scope, dev_ctx);
+      zero_op->Run(scope, dev_place);
     } else {
       auto &out_grad_tensor = out_grad_var->Get<framework::LoDTensor>();
       auto *in_grad_tensor = in_grad_var->GetMutable<framework::LoDTensor>();

paddle/operators/roi_pool_op.cc

@@ -68,7 +68,7 @@ class ROIPoolOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),
@@ -89,7 +89,7 @@ class ROIPoolGradOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),

paddle/operators/save_load_op_test.cc

@@ -21,7 +21,7 @@ USE_NO_KERNEL_OP(load);
 TEST(SaveLoadOp, CPU) {
   paddle::framework::Scope scope;
   paddle::platform::CPUPlace place;
-  paddle::platform::CPUDeviceContext ctx(place);
+
   auto var = scope.Var("test_var");
   auto tensor = var->GetMutable<paddle::framework::LoDTensor>();
   tensor->Resize({10, 10});
@@ -42,13 +42,13 @@ TEST(SaveLoadOp, CPU) {
 
   auto save_op = paddle::framework::OpRegistry::CreateOp(
       "save", {{"X", {"test_var"}}}, {}, attrs);
-  save_op->Run(scope, ctx);
+  save_op->Run(scope, place);
 
   auto load_var = scope.Var("out_var");
   auto target = load_var->GetMutable<paddle::framework::LoDTensor>();
   auto load_op = paddle::framework::OpRegistry::CreateOp(
       "load", {}, {{"Out", {"out_var"}}}, attrs);
-  load_op->Run(scope, ctx);
+  load_op->Run(scope, place);
   int* actual = target->data<int>();
   for (int64_t i = 0; i < tensor->numel(); ++i) {
     EXPECT_EQ(expect[i], actual[i]);

paddle/operators/save_op.cc

@@ -21,6 +21,7 @@
 #include "paddle/framework/framework.pb.h"
 #include "paddle/framework/lod_tensor.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/platform/device_context.h"
 
 namespace paddle {
 namespace operators {
@@ -62,7 +63,7 @@ class SaveOp : public framework::OperatorBase {
          const framework::AttributeMap &attrs)
       : OperatorBase(type, inputs, outputs, attrs) {}
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &place) const override {
     auto filename = Attr<std::string>("file_path");
     auto overwrite = Attr<bool>("overwrite");
 
@@ -88,6 +89,11 @@ class SaveOp : public framework::OperatorBase {
                    "SaveOp only support LoDTensor, %s has wrong type", iname);
 
     auto &tensor = var->Get<framework::LoDTensor>();
+
+    // get device context from pool
+    platform::DeviceContextPool &pool = platform::DeviceContextPool::Get();
+    auto &dev_ctx = *pool.Borrow(place);
+
     framework::SerializeToStream(fout, tensor, dev_ctx);
   }
 };

paddle/operators/scatter_op.cc

@@ -49,7 +49,7 @@ class ScatterOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("Ref")->type()),
@@ -68,7 +68,7 @@ class ScatterGradOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("Ref")->type()),

paddle/operators/send_op.cc

@@ -34,45 +34,56 @@ class SendOp : public framework::OperatorBase {
          const framework::AttributeMap &attrs)
       : OperatorBase(type, inputs, outputs, attrs) {
     // init client when the operator is created at runtime.
-    if (!client_) {
-      std::string endpoint = Attr<std::string>("endpoint");
-      client_.reset(new detail::RPCClient(
-          grpc::CreateChannel(endpoint, grpc::InsecureChannelCredentials())));
-      // TODO(typhoonzero): how to call InitVariables
+    std::vector<std::string> endpoints =
+        Attr<std::vector<std::string>>("endpoints");
+    for (auto ep : endpoints) {
+      client_map_[ep].reset(new detail::RPCClient(
+          grpc::CreateChannel(ep, grpc::InsecureChannelCredentials())));
     }
   }
   void Run(const framework::Scope &scope,
            const platform::DeviceContext &dev_ctx) const override {
-    auto iname = Input("X");
-    auto oname = Output("Out");
-    // TODO(typhoonzero): currently it's non-blocking,
-    // should block until server responds.
-    bool ret = client_->SendVariable(scope, iname, oname);
+    auto ins = Inputs("X");
+    std::vector<std::string> epmap = Attr<std::vector<std::string>>("epmap");
+    // TODO(typhoonzero): use async calls to send multiple variable asyncly.
+    for (size_t i = 0; i < ins.size(); ++i) {
+      bool ret = client_map_[epmap[i]]->SendVariable(scope, ins[i]);
       if (!ret) {
-      LOG(ERROR) << "send variable error";
+        LOG(ERROR) << "send variable error: " << ins[i];
+      }
+    }
+    // TODO(typhoonzero): support async optimization
+    client_map_[epmap[0]]->Wait();
+    for (size_t i = 0; i < ins.size(); ++i) {
+      bool ret = client_map_[epmap[i]]->GetVariable(scope, ins[i]);
+      if (!ret) {
+        LOG(ERROR) << "GetVariable error: " << ins[i];
+      }
     }
   }
 
  protected:
-  std::shared_ptr<detail::RPCClient> client_{nullptr};
+  mutable std::unordered_map<std::string, std::shared_ptr<detail::RPCClient>>
+      client_map_;
 };
 
 class SendOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   SendOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "(Tensor) Input tensor to be saved");
-    AddOutput("Out", "(Tensor) Output fetched from server");
+    AddInput("X", "(Tensor) Input tensor to be send").AsDuplicable();
     AddComment(R"DOC(
 Recv operator
 
 This operator will recv tensor from send_op
 )DOC");
-    AddAttr<std::string>("endpoint",
-                         "(string, default 127.0.0.1:6164)"
-                         "IP address to listen on.")
-        .SetDefault("127.0.0.1:6164")
-        .AddCustomChecker([](const std::string &ip) { return !ip.empty(); });
+    AddAttr<std::vector<std::string>>("endpoints",
+                                      "(string vector, default 127.0.0.1:6164)"
+                                      "Server endpoints to send variables to.");
+    AddAttr<std::vector<std::string>>("epmap",
+                                      "(string vector, default 127.0.0.1:6164)"
+                                      "Server endpoints in the order of input "
+                                      "variables for mapping");
   }
 };
 

paddle/operators/send_recv_op_test.cc

@@ -16,12 +16,14 @@
 // a RemoteOptimizer.
 
 #include <unistd.h>
+#include <string>
 #include <thread>
 
 #include "gtest/gtest.h"
 #include "paddle/framework/op_registry.h"
 #include "paddle/framework/operator.h"
 #include "paddle/framework/program_desc.h"
+#include "paddle/string/printf.h"
 
 USE_NO_KERNEL_OP(send);
 USE_NO_KERNEL_OP(recv);
@@ -33,30 +35,33 @@ std::unique_ptr<paddle::framework::OperatorBase> recv_op;
 void InitTensorsInScope(paddle::framework::Scope &scope,
                         paddle::platform::CPUPlace &place) {
   paddle::platform::CPUDeviceContext ctx(place);
-  auto var = scope.Var("X");
+  for (int i = 0; i < 2; ++i) {
+    auto var_name = paddle::string::Sprintf("x%d", i);
+    auto var = scope.Var(var_name);
     auto tensor = var->GetMutable<paddle::framework::LoDTensor>();
     tensor->Resize({10, 10});
     float *expect = tensor->mutable_data<float>(place);
     for (int64_t i = 0; i < tensor->numel(); ++i) {
       expect[i] = static_cast<float>(i);
     }
+  }
 
   auto out_var = scope.Var("Out");
   auto out_tensor = out_var->GetMutable<paddle::framework::LoDTensor>();
   out_tensor->Resize({10, 10});
-  tensor->mutable_data<float>(place);  // allocate
+  out_tensor->mutable_data<float>(place);  // allocate
 }
 
 void AddOp(const std::string &type,
            const paddle::framework::VariableNameMap &inputs,
            const paddle::framework::VariableNameMap &outputs,
            paddle::framework::AttributeMap attrs,
-           paddle::framework::BlockDescBind *block) {
+           paddle::framework::BlockDesc *block) {
   // insert output
   for (auto kv : outputs) {
     for (auto v : kv.second) {
       auto var = block->Var(v);
-      var->SetDataType(paddle::framework::DataType::FP32);
+      var->SetDataType(paddle::framework::proto::DataType::FP32);
     }
   }
 
@@ -78,10 +83,10 @@ void StartServerNet() {
   InitTensorsInScope(scope, place);
 
   // sub program run in recv_op, for simple test we use sum
-  paddle::framework::ProgramDescBind program;
-  paddle::framework::BlockDescBind *block = program.MutableBlock(0);
+  paddle::framework::ProgramDesc program;
+  paddle::framework::BlockDesc *block = program.MutableBlock(0);
   // X for server side tensors, RX for received tensers, must be of same shape.
-  AddOp("sum", {{"X", {"X", "RX"}}}, {{"Out", {"Out"}}}, {}, block);
+  AddOp("sum", {{"X", {"x0", "x1"}}}, {{"Out", {"Out"}}}, {}, block);
 
   paddle::framework::AttributeMap attrs;
   attrs.insert({"endpoint", std::string("127.0.0.1:6174")});
@@ -89,8 +94,8 @@ void StartServerNet() {
   PADDLE_ENFORCE(program.Proto()->SerializeToString(&program_proto));
 
   attrs.insert({"OptimizeProgram", program_proto});
-  recv_op = paddle::framework::OpRegistry::CreateOp("recv", {{"RX", {"RX"}}},
-                                                    {{"Out", {"Out"}}}, attrs);
+  recv_op = paddle::framework::OpRegistry::CreateOp(
+      "recv", {{"RX", {"x0", "x1"}}}, {{"Out", {"Out"}}}, attrs);
   paddle::platform::CPUDeviceContext ctx(place);
   recv_op->Run(scope, ctx);
 }
@@ -107,11 +112,11 @@ TEST(SendRecvOp, CPU) {
   attrs.insert({"endpoint", std::string("127.0.0.1:6174")});
 
   auto send_op = paddle::framework::OpRegistry::CreateOp(
-      "send", {{"X", {"X"}}}, {{"Out", {"Out"}}}, attrs);
+      "send", {{"X", {"x0", "x1"}}}, {{"Out", {"Out"}}}, attrs);
   paddle::platform::CPUDeviceContext ctx(place);
   send_op->Run(scope, ctx);
 
-  auto in_var = scope.Var("X");
+  auto in_var = scope.Var("x0");
   auto tensor = in_var->GetMutable<paddle::framework::LoDTensor>();
   float *expected = tensor->data<float>();
 

paddle/operators/sequence_concat_op.cc

@@ -124,8 +124,9 @@ class SequenceConcatGradOp : public framework::OperatorWithKernel {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP(sequence_concat, ops::SequenceConcatOp, ops::SequenceConcatOpMaker,
-            sequence_concat_grad, ops::SequenceConcatGradOp);
+REGISTER_OP_EX(sequence_concat, ops::SequenceConcatOp,
+               ops::SequenceConcatOpMaker, sequence_concat_grad,
+               ops::SequenceConcatGradOp, false);
 REGISTER_OP_CPU_KERNEL(
     sequence_concat,
     ops::SequenceConcatOpKernel<paddle::platform::CPUDeviceContext, float>);

paddle/operators/sequence_pool_op.cc

@@ -49,7 +49,7 @@ class SequencePoolOpMaker : public framework::OpProtoAndCheckerMaker {
         .AsIntermediate();
     AddAttr<std::string>(
         "pooltype",
-        "(int, default AVERAGE) the pooling pooltype of SequencePoolOp.")
+        "(string, default 'AVERAGE') the pooling pooltype of SequencePoolOp.")
         .SetDefault("AVERAGE")
         .InEnum({"AVERAGE", "SUM", "SQRT", "LAST", "FIRST", "MAX"});
     AddComment(R"DOC(
@@ -107,7 +107,7 @@ class SequencePoolGradOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("X")->type()),

paddle/operators/sequence_slice_op.cc

@@ -48,7 +48,7 @@ class SequenceSliceOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::LoDTensor>("X")->type()),
@@ -69,7 +69,7 @@ class SequenceSliceGradOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::LoDTensor>("X")->type()),

paddle/operators/shrink_rnn_memory_op.cc

@@ -27,11 +27,11 @@ class ShrinkRNNMemoryOp : public ArrayOp {
       : ArrayOp(type, inputs, outputs, attrs) {}
 
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &place) const override {
     auto *x_var = scope.FindVar(Input("X"));
     PADDLE_ENFORCE(x_var != nullptr, "Input X must be set");
     auto &x_tensor = x_var->Get<framework::LoDTensor>();
-    size_t offset = this->GetOffset(scope, dev_ctx);
+    size_t offset = this->GetOffset(scope, place);
     auto *rank_table_var = scope.FindVar(Input("RankTable"));
     PADDLE_ENFORCE(rank_table_var != nullptr, "RankTable must be set");
     auto &rank_table = rank_table_var->Get<framework::LoDRankTable>();
@@ -93,7 +93,7 @@ class ShrinkRNNMemoryGradOp : public ArrayOp {
       : ArrayOp(type, inputs, outputs, attrs) {}
 
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &place) const override {
     auto *dout_var = scope.FindVar(Input(framework::GradVarName("Out")));
     auto *dx_var = scope.FindVar(Output(framework::GradVarName("X")));
     PADDLE_ENFORCE(dx_var != nullptr, "Input Gradient should not be nullptr");
@@ -105,6 +105,10 @@ class ShrinkRNNMemoryGradOp : public ArrayOp {
     dx_tensor.Resize(x_tensor.dims());
     dx_tensor.mutable_data(x_tensor.place(), x_tensor.type());
 
+    // get device context from pool
+    platform::DeviceContextPool &pool = platform::DeviceContextPool::Get();
+    auto &dev_ctx = *pool.Borrow(place);
+
     if (dout_var == nullptr) {  // dx_tensor fill zero
       math::set_constant(dev_ctx, &dx_tensor, 0.0f);
     } else {

paddle/operators/softmax_with_cross_entropy_op.cc

@@ -118,7 +118,7 @@ class SoftmaxWithCrossEntropyOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("Logits")->type()),
@@ -159,7 +159,7 @@ class SoftmaxWithCrossEntropyOpGrad : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(

paddle/operators/split_lod_tensor_op.cc

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #include "paddle/framework/op_registry.h"
 #include "paddle/memory/memcpy.h"
+#include "paddle/platform/device_context.h"
 
 namespace paddle {
 namespace operators {
@@ -33,7 +34,7 @@ class SplitLoDTensorOp : public framework::OperatorBase {
                    const framework::AttributeMap &attrs)
       : OperatorBase(type, inputs, outputs, attrs) {}
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &dev_place) const override {
     auto &x = scope.FindVar(Input("X"))->Get<framework::LoDTensor>();
     auto &mask = scope.FindVar(Input("Mask"))->Get<framework::LoDTensor>();
     auto *out_true =
@@ -44,6 +45,9 @@ class SplitLoDTensorOp : public framework::OperatorBase {
     auto &x_lod = x.lod();
     auto &mask_dim = mask.dims();
 
+    platform::DeviceContextPool &pool = platform::DeviceContextPool::Get();
+    auto &dev_ctx = *pool.Borrow(dev_place);
+
     std::unique_ptr<framework::LoDTensor> cpu_mask{new framework::LoDTensor()};
     if (platform::is_cpu_place(mask.place())) {
       cpu_mask->ShareDataWith(mask);

paddle/operators/strided_memcpy_test.cc

@@ -82,11 +82,13 @@ TEST(StridedMemcpy, GPUCrop) {
   };
   // clang-format on
 
-  platform::GPUPlace gpu0(0);
+  platform::CUDAPlace gpu0(0);
   platform::CPUPlace cpu;
 
+  platform::CUDADeviceContext ctx(gpu0);
+
   int* gpu_src = reinterpret_cast<int*>(memory::Alloc(gpu0, sizeof(src)));
-  memory::Copy(gpu0, gpu_src, cpu, src, sizeof(src));
+  memory::Copy(gpu0, gpu_src, cpu, src, sizeof(src), ctx.stream());
 
   framework::DDim src_stride({5, 1});
 
@@ -96,7 +98,6 @@ TEST(StridedMemcpy, GPUCrop) {
   framework::DDim dst_dim({2, 2});
   framework::DDim dst_stride({2, 1});
 
-  platform::CUDADeviceContext ctx(gpu0);
   StridedMemcpy<int>(ctx, gpu_src + 1, src_stride, dst_dim, dst_stride,
                      gpu_dst);
 
@@ -120,11 +121,12 @@ TEST(StridedMemcpy, GPUConcat) {
   };
   // clang-format on
 
-  platform::GPUPlace gpu0(0);
+  platform::CUDAPlace gpu0(0);
   platform::CPUPlace cpu;
+  platform::CUDADeviceContext ctx(gpu0);
 
   int* gpu_src = reinterpret_cast<int*>(memory::Alloc(gpu0, sizeof(src)));
-  memory::Copy(gpu0, gpu_src, cpu, src, sizeof(src));
+  memory::Copy(gpu0, gpu_src, cpu, src, sizeof(src), ctx.stream());
 
   int dst[8];
   int* gpu_dst = reinterpret_cast<int*>(memory::Alloc(gpu0, sizeof(dst)));
@@ -132,7 +134,6 @@ TEST(StridedMemcpy, GPUConcat) {
   framework::DDim src_stride({2, 1});
   framework::DDim dst_dim({2, 2});
   framework::DDim dst_stride({4, 1});
-  platform::CUDADeviceContext ctx(gpu0);
 
   StridedMemcpy<int>(ctx, gpu_src, src_stride, dst_dim, dst_stride, gpu_dst);
   StridedMemcpy<int>(ctx, gpu_src, src_stride, dst_dim, dst_stride,

paddle/operators/sum_op.cc

@@ -53,7 +53,7 @@ class SumOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     auto x_vars = ctx.MultiInputVar("X");
     if (x_vars[0]->IsType<framework::LoDTensor>()) {
@@ -170,7 +170,7 @@ class SumGradMaker : public framework::GradOpDescMakerBase {
   using framework::GradOpDescMakerBase::GradOpDescMakerBase;
 
   std::vector<std::unique_ptr<framework::OpDesc>> operator()() const override {
-    auto x_grads = InputGrad("X");
+    auto x_grads = InputGrad("X", false);
     std::vector<std::unique_ptr<framework::OpDesc>> grad_ops;
     grad_ops.reserve(x_grads.size());
     auto og = OutputGrad("Out");

paddle/operators/tensor_array_read_write_op.cc

@@ -25,11 +25,11 @@ class WriteToArrayOp : public ArrayOp {
       : ArrayOp(type, inputs, outputs, attrs) {}
 
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &place) const override {
     auto *x = scope.FindVar(Input("X"));
     if (x == nullptr) return;
     auto &x_tensor = x->Get<framework::LoDTensor>();
-    size_t offset = GetOffset(scope, dev_ctx);
+    size_t offset = GetOffset(scope, place);
     auto *out =
         scope.FindVar(Output("Out"))->GetMutable<framework::LoDTensorArray>();
     if (offset >= out->size()) {
@@ -39,7 +39,11 @@ class WriteToArrayOp : public ArrayOp {
     }
     if (x_tensor.memory_size() > 0) {
       auto *out_tensor = &out->at(offset);
-      CopyFrom(x_tensor, dev_ctx.GetPlace(), dev_ctx, out_tensor);
+
+      platform::DeviceContextPool &pool = platform::DeviceContextPool::Get();
+      auto &dev_ctx = *pool.Borrow(place);
+
+      CopyFrom(x_tensor, place, dev_ctx, out_tensor);
       out_tensor->set_lod(x_tensor.lod());
     } else {
       VLOG(10) << "WARNING: The input tensor 'x_tensor' holds no memory, so "
@@ -119,17 +123,18 @@ class ReadFromArrayOp : public ArrayOp {
                   const framework::AttributeMap &attrs)
       : ArrayOp(type, inputs, outputs, attrs) {}
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &place) const override {
     auto *x = scope.FindVar(Input("X"));
     PADDLE_ENFORCE(x != nullptr, "X must be set");
     auto &x_array = x->Get<framework::LoDTensorArray>();
     auto *out = scope.FindVar(Output("Out"));
     PADDLE_ENFORCE(out != nullptr, "Out must be set");
     auto *out_tensor = out->GetMutable<framework::LoDTensor>();
-    size_t offset = GetOffset(scope, dev_ctx);
+    size_t offset = GetOffset(scope, place);
     if (offset < x_array.size()) {
-      framework::CopyFrom(x_array[offset], dev_ctx.GetPlace(), dev_ctx,
-                          out_tensor);
+      platform::DeviceContextPool &pool = platform::DeviceContextPool::Get();
+      auto &dev_ctx = *pool.Borrow(place);
+      framework::CopyFrom(x_array[offset], place, dev_ctx, out_tensor);
       out_tensor->set_lod(x_array[offset].lod());
     } else {
       VLOG(10) << "offset " << offset << " >= " << x_array.size();

paddle/operators/top_k_op.cu

@@ -283,7 +283,7 @@ class TopkOpCUDAKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
+                   "It must use CUDAPlace.");
     auto* input = ctx.Input<Tensor>("X");
     auto* output = ctx.Output<Tensor>("Out");
     auto* indices = ctx.Output<Tensor>("Indices");

paddle/operators/transpose_op.cc

@@ -70,16 +70,17 @@ class TransposeOpMaker : public framework::OpProtoAndCheckerMaker {
 Transpose Operator.
 
 The input tensor will be permuted according to the axis values given.
-The op functions similar to how numpy.transpose works in python.
-For example:
- >> input = numpy.arange(6).reshape((2,3))
- >> input
- array([[0, 1, 2],
+The op functions is similar to how numpy.transpose works in python.
+
+For example: input = numpy.arange(6).reshape((2,3))
+the input is:
+array([[0, 1, 2],
       [3, 4, 5]])
- >> axis = [1, 0]
- >> output = input.transpose(axis)
- >> output
- array([[0, 3],
+given axis is: [1, 0]
+
+output = input.transpose(axis)
+then the output is:
+array([[0, 3],
        [1, 4],
        [2, 5]])
 So, given a input tensor of shape(N, C, H, W) and the axis is {0, 2, 3, 1},

paddle/operators/uniform_random_op.cc

@@ -63,7 +63,7 @@ class UniformRandomOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         static_cast<framework::proto::DataType>(ctx.Attr<int>("dtype")),

paddle/operators/unpool_op.cc

@@ -53,16 +53,14 @@ class Unpool2dOpMaker : public framework::OpProtoAndCheckerMaker {
         "(string), unpooling type, can be \"max\" for max-unpooling ")
         .InEnum({"max"});
     AddComment(R"DOC(
-        "Input shape: $(N, C_{in}, H_{in}, W_{in})$
-        Output shape: $(N, C_{out}, H_{out}, W_{out})$
-        Where
-          $$
-            H_{out} = (H_{in}−1) * strides[0] − 2 * paddings[0] + ksize[0] \\
-            W_{out} = (W_{in}−1) * strides[1] − 2 * paddings[1] + ksize[1]
-          $$
-        Paper: http://www.matthewzeiler.com/wp-content/uploads/2017
-        /07/iccv2011.pdf
-        )DOC");
+Input shape is: $(N, C_{in}, H_{in}, W_{in})$, Output shape is:
+$(N, C_{out}, H_{out}, W_{out})$, where
+$$
+H_{out} = (H_{in}−1) * strides[0] − 2 * paddings[0] + ksize[0] \\
+W_{out} = (W_{in}−1) * strides[1] − 2 * paddings[1] + ksize[1]
+$$
+Paper: http://www.matthewzeiler.com/wp-content/uploads/2017/07/iccv2011.pdf
+)DOC");
   }
 };
 
@@ -73,7 +71,7 @@ int OutputSize(int input_size, int ksize, int padding, int stride) {
 
 class UnpoolOp : public framework::OperatorWithKernel {
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),
@@ -112,7 +110,7 @@ class UnpoolOp : public framework::OperatorWithKernel {
 
 class UnpoolOpGrad : public framework::OperatorWithKernel {
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),

paddle/operators/while_op.cc

@@ -40,13 +40,14 @@ class WhileOp : public framework::OperatorBase {
       : framework::OperatorBase(type, inputs, outputs, attrs) {}
 
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
+           const platform::Place &dev_place) const override {
     PADDLE_ENFORCE_NOT_NULL(scope.FindVar(Input(kCondition)));
     auto &cond = scope.FindVar(Input(kCondition))->Get<LoDTensor>();
     PADDLE_ENFORCE_EQ(cond.dims(), paddle::framework::make_ddim({1}));
 
-    framework::Executor executor(dev_ctx);
+    framework::Executor executor(dev_place);
     auto *block = Attr<framework::BlockDesc *>(kStepBlock);
+
     auto *program = block->Program();
 
     auto step_scopes =
@@ -97,8 +98,8 @@ class WhileGradOp : public framework::OperatorBase {
       : framework::OperatorBase(type, inputs, outputs, attrs) {}
 
   void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {
-    framework::Executor executor(dev_ctx);
+           const platform::Place &dev_place) const override {
+    framework::Executor executor(dev_place);
     auto *block = Attr<framework::BlockDesc *>(kStepBlock);
     auto *program = block->Program();
 
@@ -189,7 +190,7 @@ class WhileGradOp : public framework::OperatorBase {
             auto zero_op = framework::OpRegistry::CreateOp(
                 "fill_constant", framework::VariableNameMap{},
                 {{"Out", {pg_names[param_id]}}}, attrs);
-            zero_op->Run(scope, dev_ctx);
+            zero_op->Run(scope, dev_place);
           }
         }
 
@@ -197,7 +198,7 @@ class WhileGradOp : public framework::OperatorBase {
         auto sum_op = framework::OpRegistry::CreateOp(
             "sum", {{"X", {pg_names[param_id], new_inside_name}}},
             {{"Out", {pg_names[param_id]}}}, framework::AttributeMap{});
-        sum_op->Run(cur_scope, dev_ctx);
+        sum_op->Run(cur_scope, dev_place);
         cur_scope.Rename(new_inside_name, inside_grad_name);
       }
     }

paddle/platform/CMakeLists.txt

@@ -25,7 +25,7 @@ ENDIF()
 # avoiding cycle dependencies
 cc_library(device_context SRCS device_context.cc DEPS memory buddy_allocator
     system_allocator memory_block meta_data meta_cache place eigen3 ${GPU_CTX_DEPS})
-nv_test(device_context_test SRCS device_context_test.cc DEPS device_context gpu_info)
+nv_test(device_context_test SRCS device_context_test.cu 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 device_context)

paddle/platform/cuda_profiler.h

@@ -22,23 +22,7 @@ namespace paddle {
 namespace platform {
 
 void CudaProfilerInit(std::string output_file, std::string output_mode,
-                      std::vector<std::string> config_flags) {
-  std::array<char, 128> buf;
-  std::string tmpl = "/tmp/cuda_profile_config.XXXXXX";
-  PADDLE_ENFORCE_LT(tmpl.size(), buf.size());
-  memcpy(buf.data(), tmpl.data(), tmpl.size());
-  auto result = mktemp(buf.data());
-  PADDLE_ENFORCE(strlen(result) != 0);
-  std::string config_file = result;
-
-  {
-    std::ofstream ofs(config_file, std::ios::out | std::ios::trunc);
-    PADDLE_ENFORCE(ofs.is_open(), "ofstream: ", ofs.rdstate());
-    for (const auto& line : config_flags) {
-      ofs << line << std::endl;
-    }
-  }
-
+                      std::string config_file) {
   PADDLE_ENFORCE(output_mode == "kvp" || output_mode == "csv");
   cudaOutputMode_t mode = output_mode == "csv" ? cudaCSV : cudaKeyValuePair;
   PADDLE_ENFORCE(

paddle/platform/device_context.cc

@@ -15,6 +15,59 @@ limitations under the License. */
 namespace paddle {
 namespace platform {
 
+DeviceContextPool* DeviceContextPool::pool = nullptr;
+
+const platform::DeviceContext* DeviceContextPool::Borrow(
+    const platform::Place& place) {
+  auto it = device_contexts_.find(place);
+  if (it == device_contexts_.end()) {
+    PADDLE_THROW(
+        "'Place' is not supported, Please re-compile with WITH_GPU "
+        "option");
+  }
+  return it->second;
+}
+
+std::vector<const platform::DeviceContext*> DeviceContextPool::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 it = device_contexts_.find(place);
+    if (it != device_contexts_.end()) {
+      borrowed_contexts.emplace_back(it->second);
+    } else {
+      PADDLE_THROW(
+          "'Place' is not supported, Please re-compile with WITH_GPU "
+          "option");
+    }
+  }
+  return borrowed_contexts;
+}
+
+DeviceContextPool::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::CUDAPlace>(places[i])));
+#else
+      PADDLE_THROW(
+          "'CUDAPlace' is not supported, Please re-compile with WITH_GPU "
+          "option");
+#endif
+    }
+  }
+}
+
 CPUDeviceContext::CPUDeviceContext() {
   eigen_device_.reset(new Eigen::DefaultDevice());
 }
@@ -38,7 +91,7 @@ class EigenCudaStreamDevice : public Eigen::StreamInterface {
   }
   ~EigenCudaStreamDevice() override {}
 
-  void Reinitialize(const cudaStream_t* cuda_stream, GPUPlace place) {
+  void Reinitialize(const cudaStream_t* cuda_stream, CUDAPlace place) {
     stream_ = cuda_stream;
     place_ = place;
     device_prop_ = &Eigen::m_deviceProperties[place.device];
@@ -77,14 +130,14 @@ class EigenCudaStreamDevice : public Eigen::StreamInterface {
   }
 
  private:
-  GPUPlace place_;
+  CUDAPlace place_;
   const cudaStream_t* stream_;         // not owned;
   const cudaDeviceProp* device_prop_;  // not owned;
   mutable void* scratch_;
   mutable unsigned int* semaphore_;
 };
 
-CUDADeviceContext::CUDADeviceContext(GPUPlace place) : place_(place) {
+CUDADeviceContext::CUDADeviceContext(CUDAPlace place) : place_(place) {
   SetDeviceId(place_.device);
   PADDLE_ENFORCE(cudaStreamCreate(&stream_));
   eigen_stream_.reset(new EigenCudaStreamDevice());
@@ -125,20 +178,18 @@ cudnnHandle_t CUDADeviceContext::cudnn_handle() const { return cudnn_handle_; }
 
 cudaStream_t CUDADeviceContext::stream() const { return stream_; }
 
-CUDNNDeviceContext::CUDNNDeviceContext(CUDNNPlace place)
-    : CUDADeviceContext(place), place_(place) {
+CUDNNDeviceContext::CUDNNDeviceContext(CUDAPlace place)
+    : CUDADeviceContext(place) {
   PADDLE_ENFORCE(dynload::cudnnCreate(&cudnn_handle_));
   PADDLE_ENFORCE(dynload::cudnnSetStream(cudnn_handle_, stream()));
 }
 
 CUDNNDeviceContext::~CUDNNDeviceContext() {
-  SetDeviceId(place_.device);
+  SetDeviceId(boost::get<CUDAPlace>(GetPlace()).device);
   Wait();
   PADDLE_ENFORCE(dynload::cudnnDestroy(cudnn_handle_));
 }
 
-Place CUDNNDeviceContext::GetPlace() const { return CUDNNPlace(); }
-
 cudnnHandle_t CUDNNDeviceContext::cudnn_handle() const { return cudnn_handle_; }
 
 #endif

paddle/platform/device_context.h

@@ -11,8 +11,8 @@ limitations under the License. */
 
 #pragma once
 
-#include "paddle/platform/enforce.h"
-#include "paddle/platform/place.h"
+#include <memory>
+#include <unordered_map>
 
 #ifdef PADDLE_WITH_CUDA
 #include "paddle/platform/dynload/cublas.h"
@@ -20,10 +20,13 @@ limitations under the License. */
 #include "paddle/platform/gpu_info.h"
 #define EIGEN_USE_GPU
 #endif
-#include <memory>
+
+#include "paddle/platform/enforce.h"
 #include "paddle/platform/place.h"
 #include "unsupported/Eigen/CXX11/Tensor"
 
+#include "glog/logging.h"
+
 namespace paddle {
 namespace platform {
 
@@ -55,7 +58,7 @@ class EigenCudaStreamDevice;
 
 class CUDADeviceContext : public DeviceContext {
  public:
-  explicit CUDADeviceContext(GPUPlace place);
+  explicit CUDADeviceContext(CUDAPlace place);
   virtual ~CUDADeviceContext();
 
   /*! \brief  Wait for all operations completion in the stream. */
@@ -77,7 +80,7 @@ class CUDADeviceContext : public DeviceContext {
   cudaStream_t stream() const;
 
  private:
-  GPUPlace place_;
+  CUDAPlace place_;
 
   std::unique_ptr<Eigen::GpuDevice> eigen_device_;
   std::unique_ptr<EigenCudaStreamDevice> eigen_stream_;
@@ -89,21 +92,63 @@ class CUDADeviceContext : public DeviceContext {
 
 class CUDNNDeviceContext : public CUDADeviceContext {
  public:
-  explicit CUDNNDeviceContext(CUDNNPlace place);
+  explicit CUDNNDeviceContext(CUDAPlace place);
   virtual ~CUDNNDeviceContext();
 
-  /*! \brief  Return place in the device context. */
-  Place GetPlace() const final;
-
   /*! \brief  Return cudnn  handle in the device context. */
   cudnnHandle_t cudnn_handle() const;
 
  private:
   cudnnHandle_t cudnn_handle_;
-  CUDNNPlace place_;
 };
 
 #endif
 
+/*! \brief device context pool singleton */
+class DeviceContextPool {
+ public:
+  explicit DeviceContextPool(const std::vector<platform::Place>& places);
+
+  static DeviceContextPool& Get() {
+    PADDLE_ENFORCE_NOT_NULL(pool, "Need to Create DeviceContextPool first!");
+    return *pool;
+  }
+
+  /*! \brief  Create should only called by Init function */
+  static DeviceContextPool& Create(const std::vector<platform::Place>& places) {
+    if (pool == nullptr) {
+      pool = new DeviceContextPool(places);
+    }
+    return *pool;
+  }
+
+  /*! \brief  Return handle of single device context. */
+  const platform::DeviceContext* Borrow(const platform::Place& place);
+
+  /*! \brief  Return handle of multi-device context. */
+  std::vector<const platform::DeviceContext*> Borrow(
+      const std::vector<platform::Place>& places);
+
+  ~DeviceContextPool() {}
+
+ private:
+  static DeviceContextPool* pool;
+  constexpr static int LEFT_SHIFT = 8;
+  struct Hash {
+    std::hash<int> hash_;
+    size_t operator()(const platform::Place& place) const {
+      int pre_hash = place.which() + (1 << LEFT_SHIFT);
+      if (platform::is_gpu_place(place)) {
+        pre_hash += boost::get<platform::CUDAPlace>(place).GetDeviceId();
+      }
+      return hash_(pre_hash);
+    }
+  };
+  std::unordered_map<const platform::Place, const platform::DeviceContext*,
+                     Hash>
+      device_contexts_;
+  DISABLE_COPY_AND_ASSIGN(DeviceContextPool);
+};
+
 }  // namespace platform
 }  // namespace paddle

paddle/platform/device_context_test.cc → paddle/platform/device_context_test.cu

@@ -12,17 +12,19 @@ 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/platform/device_context.h"
 #include "gtest/gtest.h"
+#include "paddle/platform/device_context.h"
+
+#include "glog/logging.h"
 
 TEST(Device, Init) {
   using paddle::platform::DeviceContext;
   using paddle::platform::CUDADeviceContext;
-  using paddle::platform::GPUPlace;
+  using paddle::platform::CUDAPlace;
 
   int count = paddle::platform::GetCUDADeviceCount();
   for (int i = 0; i < count; i++) {
-    CUDADeviceContext* device_context = new CUDADeviceContext(GPUPlace(i));
+    CUDADeviceContext* device_context = new CUDADeviceContext(CUDAPlace(i));
     Eigen::GpuDevice* gpu_device = device_context->eigen_device();
     ASSERT_NE(nullptr, gpu_device);
     delete device_context;
@@ -31,11 +33,11 @@ TEST(Device, Init) {
 
 TEST(Device, CUDADeviceContext) {
   using paddle::platform::CUDADeviceContext;
-  using paddle::platform::GPUPlace;
+  using paddle::platform::CUDAPlace;
 
   int count = paddle::platform::GetCUDADeviceCount();
   for (int i = 0; i < count; i++) {
-    CUDADeviceContext* device_context = new CUDADeviceContext(GPUPlace(i));
+    CUDADeviceContext* device_context = new CUDADeviceContext(CUDAPlace(i));
     Eigen::GpuDevice* gpu_device = device_context->eigen_device();
     ASSERT_NE(nullptr, gpu_device);
     cudnnHandle_t cudnn_handle = device_context->cudnn_handle();
@@ -49,12 +51,11 @@ TEST(Device, CUDADeviceContext) {
 
 TEST(Device, CUDNNDeviceContext) {
   using paddle::platform::CUDNNDeviceContext;
-  using paddle::platform::CUDNNPlace;
+  using paddle::platform::CUDAPlace;
   if (paddle::platform::dynload::HasCUDNN()) {
     int count = paddle::platform::GetCUDADeviceCount();
     for (int i = 0; i < count; ++i) {
-      CUDNNDeviceContext* device_context =
-          new CUDNNDeviceContext(CUDNNPlace(i));
+      CUDNNDeviceContext* device_context = new CUDNNDeviceContext(CUDAPlace(i));
       cudnnHandle_t cudnn_handle = device_context->cudnn_handle();
       ASSERT_NE(nullptr, cudnn_handle);
       ASSERT_NE(nullptr, device_context->stream());
@@ -62,3 +63,54 @@ TEST(Device, CUDNNDeviceContext) {
     }
   }
 }
+
+TEST(Device, DeviceContextPool) {
+  using paddle::platform::DeviceContextPool;
+  using paddle::platform::CUDADeviceContext;
+  using paddle::platform::Place;
+  using paddle::platform::CPUPlace;
+  using paddle::platform::CUDAPlace;
+
+  DeviceContextPool& pool = DeviceContextPool::Get();
+  auto cpu_dev_ctx1 = pool.Borrow(CPUPlace());
+  auto cpu_dev_ctx2 = pool.Borrow(CPUPlace());
+  EXPECT_TRUE(cpu_dev_ctx2 == cpu_dev_ctx1);
+
+  std::vector<Place> gpu_places;
+  int count = paddle::platform::GetCUDADeviceCount();
+  for (int i = 0; i < count; ++i) {
+    gpu_places.emplace_back(CUDAPlace(i));
+  }
+  auto dev_ctxs = pool.Borrow(gpu_places);
+  for (size_t i = 0; i < dev_ctxs.size(); ++i) {
+    auto* dev_ctx = static_cast<const CUDADeviceContext*>(dev_ctxs[i]);
+
+    // check same as CUDAPlace(i)
+    CUDAPlace place = boost::get<CUDAPlace>(dev_ctx->GetPlace());
+    EXPECT_EQ(place.GetDeviceId(), static_cast<int>(i));
+  }
+}
+
+int main(int argc, char** argv) {
+  int dev_count = paddle::platform::GetCUDADeviceCount();
+  if (dev_count <= 1) {
+    LOG(WARNING) << "Cannot test multi-gpu DeviceContextPool, because the CUDA "
+                    "device count is "
+                 << dev_count;
+    return 0;
+  }
+
+  std::vector<paddle::platform::Place> places;
+
+  places.emplace_back(paddle::platform::CPUPlace());
+  int count = paddle::platform::GetCUDADeviceCount();
+  for (int i = 0; i < count; ++i) {
+    places.emplace_back(paddle::platform::CUDAPlace(i));
+  }
+
+  VLOG(0) << " DeviceCount " << count;
+  paddle::platform::DeviceContextPool::Create(places);
+
+  testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}

paddle/platform/dynload/nccl.h

@@ -63,6 +63,8 @@ extern void LoadNCCLDSO();
   __macro(ncclAllReduce);               \
   __macro(ncclBcast);                   \
   __macro(ncclAllGather);               \
+  __macro(ncclGroupStart);              \
+  __macro(ncclGroupEnd);                \
   __macro(ncclReduce);                  \
   __macro(ncclGetErrorString);
 

paddle/platform/enforce.h

@@ -22,6 +22,7 @@ limitations under the License. */
 #include <stdexcept>
 #include <string>
 
+#include "paddle/platform/macros.h"
 #include "paddle/string/printf.h"
 #include "paddle/string/to_string.h"
 

paddle/platform/for_range.h

+/* 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. */
+
+#pragma once
+#include "paddle/platform/device_context.h"
+
+namespace paddle {
+namespace platform {
+
+template <typename DeviceContext>
+struct ForRange {
+  ForRange(const DeviceContext& dev_ctx, size_t limit);
+
+  template <typename Function>
+  void operator()(Function func) const;
+};
+
+template <>
+struct ForRange<CPUDeviceContext> {
+  ForRange(const CPUDeviceContext& dev_ctx, size_t limit) : limit_(limit) {}
+
+  template <typename Function>
+  void operator()(Function func) const {
+    for (size_t i = 0; i < limit_; ++i) {
+      func(i);
+    }
+  }
+
+  size_t limit_;
+};
+
+#ifdef __NVCC__
+template <typename Function>
+__global__ static void ForRangeElemwiseOpGridIsOne(Function func) {
+  size_t idx = static_cast<size_t>(threadIdx.x);
+  func(idx);
+}
+
+template <typename Function>
+__global__ static void ForRangeElemwiseOp(Function func, int limit) {
+  size_t idx = static_cast<size_t>(blockIdx.x * blockDim.x + threadIdx.x);
+  if (idx < limit) {
+    func(idx);
+  }
+}
+
+template <>
+struct ForRange<CUDADeviceContext> {
+  ForRange(const CUDADeviceContext& dev_ctx, size_t limit)
+      : dev_ctx_(dev_ctx), limit_(static_cast<int>(limit)) {}
+
+  template <typename Function>
+  inline void operator()(Function func) const {
+    constexpr size_t num_threads = 1024;
+    int block_size = limit_ <= num_threads ? limit_ : num_threads;
+    int grid_size = (limit_ + num_threads - 1) / num_threads;
+
+    if (grid_size == 1) {
+      ForRangeElemwiseOpGridIsOne<<<1, block_size, 0, dev_ctx_.stream()>>>(
+          func);
+    } else {
+      ForRangeElemwiseOp<<<grid_size, block_size, 0, dev_ctx_.stream()>>>(
+          func, limit_);
+    }
+  }
+
+  const CUDADeviceContext& dev_ctx_;
+  int limit_;
+};
+
+#endif
+
+}  // namespace platform
+}  // namespace paddle

paddle/platform/gpu_info.cc

@@ -97,17 +97,6 @@ void GpuMemcpyAsync(void *dst, const void *src, size_t count,
                  "cudaMemcpyAsync failed in paddle::platform::GpuMemcpyAsync");
 }
 
-void GpuMemcpySync(void *dst, const void *src, size_t count,
-                   enum cudaMemcpyKind kind) {
-  PADDLE_ENFORCE(cudaMemcpy(dst, src, count, kind),
-                 "cudaMemcpy failed in paddle::platform::GpuMemcpySync");
-  // note: cudaMemcpy may actually be asynchronous with respect to the caller,
-  //       block on stream 0 to make sure the copy has completed
-  PADDLE_ENFORCE(
-      cudaStreamSynchronize(0),
-      "cudaStreamSynchronize failed in paddle::platform::GpuMemcpySync");
-}
-
 void GpuMemcpyPeer(void *dst, int dst_device, const void *src, int src_device,
                    size_t count, cudaStream_t stream) {
   PADDLE_ENFORCE(

paddle/platform/gpu_info.h

@@ -52,10 +52,6 @@ size_t GpuMaxChunkSize();
 void GpuMemcpyAsync(void *dst, const void *src, size_t count,
                     enum cudaMemcpyKind kind, cudaStream_t stream);
 
-//! Copy memory from address src to dst synchronously.
-void GpuMemcpySync(void *dst, const void *src, size_t count,
-                   enum cudaMemcpyKind kind);
-
 //! Copy memory from one device to another device.
 void GpuMemcpyPeer(void *dst, int dst_device, const void *src, int src_device,
                    size_t count, cudaStream_t stream);

paddle/platform/nccl_test.cu

@@ -12,17 +12,19 @@
    See the License for the specific language governing permissions and
    limitations under the License. */
 
+#include <thrust/device_vector.h>
+#include <memory>
+#include <vector>
+
 #include "glog/logging.h"
 #include "gtest/gtest.h"
+
+#include "paddle/framework/init.h"
 #include "paddle/platform/device_context.h"
 #include "paddle/platform/dynload/nccl.h"
 #include "paddle/platform/enforce.h"
 #include "paddle/platform/gpu_info.h"
 
-#include <thrust/device_vector.h>
-#include <memory>
-#include <vector>
-
 static int dev_count = 0;
 
 namespace paddle {
@@ -31,7 +33,8 @@ namespace platform {
 TEST(NCCL, init) {
   std::vector<ncclComm_t> comms;
   comms.resize(dev_count);
-  dynload::ncclCommInitAll(comms.data(), dev_count, nullptr);
+  PADDLE_ENFORCE(dynload::ncclCommInitAll(comms.data(), dev_count, nullptr));
+
   for (int i = 0; i < dev_count; ++i) {
     dynload::ncclCommDestroy(comms[i]);
   }
@@ -47,7 +50,7 @@ struct PerThreadData {
 
   T* RecvBuff() { return thrust::raw_pointer_cast(recv_buff.data()); }
 
-  PerThreadData(int gpu_id, size_t size) : dev_ctx(GPUPlace(gpu_id)) {
+  PerThreadData(int gpu_id, size_t size) : dev_ctx(CUDAPlace(gpu_id)) {
     send_buff.resize(size);
     for (size_t i = 0; i < size; ++i) {
       send_buff[i] = static_cast<T>(i);
@@ -131,6 +134,18 @@ int main(int argc, char** argv) {
         << dev_count;
     return 0;
   }
+
+  std::vector<paddle::platform::Place> places;
+
+  places.emplace_back(paddle::platform::CPUPlace());
+  int count = paddle::platform::GetCUDADeviceCount();
+  for (int i = 0; i < count; ++i) {
+    places.emplace_back(paddle::platform::CUDAPlace(i));
+  }
+
+  VLOG(0) << " DeviceCount " << count;
+  paddle::platform::DeviceContextPool::Create(places);
+
   testing::InitGoogleTest(&argc, argv);
   return RUN_ALL_TESTS();
 }

paddle/platform/place.cc

@@ -23,8 +23,9 @@ class PlacePrinter : public boost::static_visitor<> {
  public:
   explicit PlacePrinter(std::ostream &os) : os_(os) {}
   void operator()(const CPUPlace &) { os_ << "CPUPlace"; }
-  void operator()(const MKLDNNPlace &) { os_ << "MKLDNNPlace"; }
-  void operator()(const GPUPlace &p) { os_ << "GPUPlace(" << p.device << ")"; }
+  void operator()(const CUDAPlace &p) {
+    os_ << "CUDAPlace(" << p.device << ")";
+  }
 
  private:
   std::ostream &os_;
@@ -37,20 +38,14 @@ static Place the_default_place;
 void set_place(const Place &place) { the_default_place = place; }
 const Place &get_place() { return the_default_place; }
 
-const GPUPlace default_gpu() { return GPUPlace(0); }
+const CUDAPlace default_gpu() { return CUDAPlace(0); }
 const CPUPlace default_cpu() { return CPUPlace(); }
-const MKLDNNPlace default_mkldnn() { return MKLDNNPlace(); }
 
 bool is_gpu_place(const Place &p) {
-  return boost::apply_visitor(IsGPUPlace(), p);
-}
-bool is_cpu_place(const Place &p) {
-  return !is_gpu_place(p) && !is_mkldnn_place(p);
+  return boost::apply_visitor(IsCUDAPlace(), p);
 }
 
-bool is_mkldnn_place(const Place &p) {
-  return boost::apply_visitor(IsMKLDNNPlace(), p);
-}
+bool is_cpu_place(const Place &p) { return !is_gpu_place(p); }
 
 bool places_are_same_class(const Place &p1, const Place &p2) {
   return p1.which() == p2.which();

paddle/platform/place.h

@@ -31,65 +31,35 @@ struct CPUPlace {
   inline bool operator!=(const CPUPlace &) const { return false; }
 };
 
-struct MKLDNNPlace {
-  MKLDNNPlace() {}
-
-  // needed for variant equality comparison
-  inline bool operator==(const MKLDNNPlace &) const { return true; }
-  inline bool operator!=(const MKLDNNPlace &) const { return false; }
-};
-
-struct GPUPlace {
-  GPUPlace() : GPUPlace(0) {}
-  explicit GPUPlace(int d) : device(d) {}
+struct CUDAPlace {
+  CUDAPlace() : CUDAPlace(0) {}
+  explicit CUDAPlace(int d) : device(d) {}
 
   inline int GetDeviceId() const { return device; }
   // needed for variant equality comparison
-  inline bool operator==(const GPUPlace &o) const { return device == o.device; }
-  inline bool operator!=(const GPUPlace &o) const { return !(*this == o); }
+  inline bool operator==(const CUDAPlace &o) const {
+    return device == o.device;
+  }
+  inline bool operator!=(const CUDAPlace &o) const { return !(*this == o); }
 
   int device;
 };
 
-struct CUDNNPlace : public GPUPlace {
-  CUDNNPlace() : GPUPlace() {}
-  explicit CUDNNPlace(int d) : GPUPlace(d) {}
-};
-
-struct IsGPUPlace : public boost::static_visitor<bool> {
+struct IsCUDAPlace : public boost::static_visitor<bool> {
   bool operator()(const CPUPlace &) const { return false; }
-  bool operator()(const MKLDNNPlace &) const { return false; }
-  bool operator()(const GPUPlace &gpu) const { return true; }
+  bool operator()(const CUDAPlace &gpu) const { return true; }
 };
 
-struct IsMKLDNNPlace : public boost::static_visitor<bool> {
-  bool operator()(const MKLDNNPlace &) const { return true; }
-  bool operator()(const CPUPlace &) const { return false; }
-  bool operator()(const GPUPlace &) const { return false; }
-};
-
-// Define the max number of Place in bit length. i.e., the max number of places
-// should be less equal than 2^(NUM_PLACE_TYPE_LIMIT_IN_BIT)
-#define NUM_PLACE_TYPE_LIMIT_IN_BIT 4
-
-typedef boost::variant<CUDNNPlace, GPUPlace, CPUPlace, MKLDNNPlace> Place;
-
-// static check number of place types is less equal than
-// 2^(NUM_PLACE_TYPE_LIMIT_IN_BIT)
-BOOST_MPL_ASSERT((boost::mpl::less_equal<
-                  Place::types::size,
-                  boost::mpl::long_<1 << NUM_PLACE_TYPE_LIMIT_IN_BIT>>));
+typedef boost::variant<CUDAPlace, CPUPlace> Place;
 
 void set_place(const Place &);
 const Place &get_place();
 
-const GPUPlace default_gpu();
+const CUDAPlace default_gpu();
 const CPUPlace default_cpu();
-const MKLDNNPlace default_mkldnn();
 
 bool is_gpu_place(const Place &);
 bool is_cpu_place(const Place &);
-bool is_mkldnn_place(const Place &);
 bool places_are_same_class(const Place &, const Place &);
 
 std::ostream &operator<<(std::ostream &, const Place &);

paddle/platform/place_test.cc

@@ -4,45 +4,34 @@
 
 TEST(Place, Equality) {
   paddle::platform::CPUPlace cpu;
-  paddle::platform::GPUPlace g0(0), g1(1), gg0(0);
-  paddle::platform::CUDNNPlace d0(0), d1(1), dd0(0);
+  paddle::platform::CUDAPlace g0(0), g1(1), gg0(0);
 
   EXPECT_EQ(cpu, cpu);
   EXPECT_EQ(g0, g0);
   EXPECT_EQ(g1, g1);
   EXPECT_EQ(g0, gg0);
-  EXPECT_EQ(d0, dd0);
 
   EXPECT_NE(g0, g1);
-  EXPECT_NE(d0, d1);
 
   EXPECT_TRUE(paddle::platform::places_are_same_class(g0, gg0));
   EXPECT_FALSE(paddle::platform::places_are_same_class(g0, cpu));
-
-  EXPECT_TRUE(paddle::platform::is_gpu_place(d0));
-  EXPECT_FALSE(paddle::platform::places_are_same_class(g0, d0));
 }
 
 TEST(Place, Default) {
   EXPECT_TRUE(paddle::platform::is_gpu_place(paddle::platform::get_place()));
   EXPECT_TRUE(paddle::platform::is_gpu_place(paddle::platform::default_gpu()));
   EXPECT_TRUE(paddle::platform::is_cpu_place(paddle::platform::default_cpu()));
-  EXPECT_TRUE(
-      paddle::platform::is_mkldnn_place(paddle::platform::default_mkldnn()));
 
+  EXPECT_FALSE(paddle::platform::is_cpu_place(paddle::platform::get_place()));
   paddle::platform::set_place(paddle::platform::CPUPlace());
   EXPECT_TRUE(paddle::platform::is_cpu_place(paddle::platform::get_place()));
-
-  paddle::platform::set_place(paddle::platform::MKLDNNPlace());
-  EXPECT_FALSE(paddle::platform::is_cpu_place(paddle::platform::get_place()));
-  EXPECT_TRUE(paddle::platform::is_mkldnn_place(paddle::platform::get_place()));
 }
 
 TEST(Place, Print) {
   {
     std::stringstream ss;
-    ss << paddle::platform::GPUPlace(1);
-    EXPECT_EQ("GPUPlace(1)", ss.str());
+    ss << paddle::platform::CUDAPlace(1);
+    EXPECT_EQ("CUDAPlace(1)", ss.str());
   }
   {
     std::stringstream ss;

paddle/platform/transform_test.cu

@@ -49,15 +49,15 @@ TEST(Transform, CPUUnary) {
 TEST(Transform, GPUUnary) {
   using namespace paddle::platform;
   using namespace paddle::memory;
-  GPUPlace gpu0(0);
+  CUDAPlace 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));
+  Copy(gpu0, gpu_buf, CPUPlace(), cpu_buf, sizeof(cpu_buf), ctx.stream());
   Transform<paddle::platform::CUDADeviceContext> trans;
   trans(ctx, gpu_buf, gpu_buf + 4, gpu_buf, Scale<float>(10));
   ctx.Wait();
-  Copy(CPUPlace(), cpu_buf, gpu0, gpu_buf, sizeof(cpu_buf));
+  Copy(CPUPlace(), cpu_buf, gpu0, gpu_buf, sizeof(cpu_buf), ctx.stream());
   Free(gpu0, gpu_buf);
   for (int i = 0; i < 4; ++i) {
     ASSERT_NEAR(cpu_buf[i], static_cast<float>(i + 1), 1e-5);
@@ -80,14 +80,14 @@ TEST(Transform, GPUBinary) {
   using namespace paddle::platform;
   using namespace paddle::memory;
   int buf[4] = {1, 2, 3, 4};
-  GPUPlace gpu0(0);
+  CUDAPlace gpu0(0);
   CUDADeviceContext ctx(gpu0);
   int* gpu_buf = static_cast<int*>(Alloc(gpu0, sizeof(buf)));
-  Copy(gpu0, gpu_buf, CPUPlace(), buf, sizeof(buf));
+  Copy(gpu0, gpu_buf, CPUPlace(), buf, sizeof(buf), ctx.stream());
   Transform<paddle::platform::CUDADeviceContext> trans;
   trans(ctx, gpu_buf, gpu_buf + 4, gpu_buf, gpu_buf, Multiply<int>());
   ctx.Wait();
-  Copy(CPUPlace(), buf, gpu0, gpu_buf, sizeof(buf));
+  Copy(CPUPlace(), buf, gpu0, gpu_buf, sizeof(buf), ctx.stream());
   Free(gpu0, gpu_buf);
   for (int i = 0; i < 4; ++i) {
     ASSERT_EQ((i + 1) * (i + 1), buf[i]);

paddle/pybind/const_value.cc

@@ -23,6 +23,11 @@ void BindConstValue(pybind11::module& m) {
   m.def("kTempVarName", [] { return framework::kTempVarName; });
   m.def("kGradVarSuffix", [] { return framework::kGradVarSuffix; });
   m.def("kZeroVarSuffix", [] { return framework::kZeroVarSuffix; });
+
+  // for kernel_hint key
+  m.def("kUseCPU", [] { return framework::kUseCPU; });
+  m.def("kUseCUDNN", [] { return framework::kUseCUDNN; });
+  m.def("kUseMKLDNN", [] { return framework::kUseMKLDNN; });
 }
 
 }  // namespace pybind

paddle/pybind/protobuf.cc

@@ -159,6 +159,7 @@ void BindBlockDesc(py::module &m) {
            py::return_value_policy::reference)
       .def("prepend_op", &BlockDesc::PrependOp,
            py::return_value_policy::reference)
+      .def("remove_op", &BlockDesc::RemoveOp)
       .def("var",
            [](BlockDesc &self, py::bytes byte_name) {
              std::string name = byte_name;
@@ -249,6 +250,12 @@ void BindOpDesc(py::module &m) {
       .def("set_attr", &OpDesc::SetAttr)
       .def("attr", &OpDesc::GetAttr)
       .def("set_block_attr", &OpDesc::SetBlockAttr)
+      .def("set_serialized_attr",
+           [](OpDesc &self, const std::string &name,
+              const py::bytes &seriralized) {
+             std::string ser(seriralized);
+             self.SetAttr(name, ser);
+           })
       .def("block_attr", &OpDesc::GetBlockAttr)
       .def("check_attrs", &OpDesc::CheckAttrs)
       .def("infer_shape", &OpDesc::InferShape)

paddle/pybind/pybind.cc

@@ -79,7 +79,7 @@ PYBIND11_PLUGIN(core) {
              self.Resize(make_ddim(dim));
            })
       .def("alloc_float",
-           [](Tensor &self, paddle::platform::GPUPlace &place) {
+           [](Tensor &self, paddle::platform::CUDAPlace &place) {
              self.mutable_data<float>(place);
            })
       .def("alloc_float",
@@ -91,7 +91,7 @@ PYBIND11_PLUGIN(core) {
              self.mutable_data<int>(place);
            })
       .def("alloc_int",
-           [](Tensor &self, paddle::platform::GPUPlace &place) {
+           [](Tensor &self, paddle::platform::CUDAPlace &place) {
              self.mutable_data<int>(place);
            })
       .def("set", PyCPUTensorSetFromArray<float>)
@@ -310,10 +310,10 @@ All parameter, weight, gradient are variables in Paddle.
                     return new paddle::platform::CPUDeviceContext();
                   })
       .def_static("create",
-                  [](paddle::platform::GPUPlace& place)
+                  [](paddle::platform::CUDAPlace& place)
                       -> paddle::platform::DeviceContext* {
 #ifndef PADDLE_WITH_CUDA
-                    PADDLE_THROW("GPUPlace is not supported in CPU device.");
+                    PADDLE_THROW("CUDAPlace is not supported in CPU device.");
 #else
                     return new paddle::platform::CUDADeviceContext(place);
 #endif
@@ -323,9 +323,9 @@ All parameter, weight, gradient are variables in Paddle.
 #ifdef PADDLE_WITH_CUDA
   py::class_<platform::Communicator>(m, "Communicator").def(py::init<>());
 #endif
-  py::class_<platform::GPUPlace>(m, "GPUPlace")
+  py::class_<platform::CUDAPlace>(m, "CUDAPlace")
       .def(py::init<int>())
-      .def("__str__", string::to_string<const platform::GPUPlace &>);
+      .def("__str__", string::to_string<const platform::CUDAPlace &>);
 
   py::class_<paddle::platform::CPUPlace>(m, "CPUPlace")
       .def(py::init<>())
@@ -338,7 +338,7 @@ All parameter, weight, gradient are variables in Paddle.
              self = cpu_place;
            })
       .def("set_place",
-           [](platform::Place &self, const platform::GPUPlace &gpu_place) {
+           [](platform::Place &self, const platform::CUDAPlace &gpu_place) {
              self = gpu_place;
            });
 
@@ -360,10 +360,10 @@ All parameter, weight, gradient are variables in Paddle.
            })
       .def("run",
            [](OperatorBase &self, const Scope &scope,
-              const platform::DeviceContext &dev_ctx) {
-             self.Run(scope, dev_ctx);
-             dev_ctx.Wait();
-           })
+              const platform::CPUPlace &place) { self.Run(scope, place); })
+      .def("run",
+           [](OperatorBase &self, const Scope &scope,
+              const platform::CUDAPlace &place) { self.Run(scope, place); })
       .def("type",
            [](const OperatorBase &op) -> std::string { return op.Type(); })
       .def("outputs",
@@ -417,7 +417,7 @@ All parameter, weight, gradient are variables in Paddle.
            });
 
   py::class_<framework::Executor>(m, "Executor")
-      .def(py::init<std::vector<platform::Place> &>())
+      .def(py::init<const platform::Place &>())
       .def("run", &Executor::Run);
 
   m.def("unique_integer", UniqueIntegerGenerator);

paddle/pybind/tensor_py.h

@@ -16,6 +16,7 @@
 #include <string>
 #include "paddle/framework/tensor.h"
 #include "paddle/memory/memcpy.h"
+#include "paddle/platform/device_context.h"
 #include "pybind11/numpy.h"
 #include "pybind11/pybind11.h"
 
@@ -61,13 +62,16 @@ struct CastToPyBufferImpl<true, I, ARGS...> {
         auto *src_ptr = static_cast<const void *>(tensor.data<CUR_TYPE>());
         auto *dst_ptr = static_cast<void *>(dst_tensor.mutable_data<CUR_TYPE>(
             tensor.dims(), platform::CPUPlace()));
-        // TODO(qijun): Here we use default CUDA stream to set GPU Tensor to
-        // a Python numpy array. It's better to manage CDUA stream unifiedly.
-        paddle::platform::GpuMemcpySync(dst_ptr, src_ptr,
-                                        sizeof(CUR_TYPE) * tensor.numel(),
-                                        cudaMemcpyDeviceToHost);
+
+        platform::DeviceContextPool &pool = platform::DeviceContextPool::Get();
+        auto dev_ctx = static_cast<const platform::CUDADeviceContext *>(
+            pool.Borrow(tensor.place()));
+
+        paddle::platform::GpuMemcpyAsync(
+            dst_ptr, src_ptr, sizeof(CUR_TYPE) * tensor.numel(),
+            cudaMemcpyDeviceToHost, dev_ctx->stream());
 #else
-        PADDLE_THROW("'GPUPlace' is not supported in CPU only device.");
+        PADDLE_THROW("'CUDAPlace' is not supported in CPU only device.");
 #endif
       } else if (paddle::platform::is_cpu_place(tensor.place())) {
         dst_tensor = tensor;
@@ -123,7 +127,7 @@ template <typename T>
 void PyCUDATensorSetFromArray(
     framework::Tensor &self,
     py::array_t<T, py::array::c_style | py::array::forcecast> array,
-    paddle::platform::GPUPlace &place) {
+    paddle::platform::CUDAPlace &place) {
   std::vector<int64_t> dims;
   dims.reserve(array.ndim());
   for (size_t i = 0; i < array.ndim(); ++i) {
@@ -132,10 +136,12 @@ void PyCUDATensorSetFromArray(
 
   self.Resize(framework::make_ddim(dims));
   auto *dst = self.mutable_data<T>(place);
-  // TODO(qijun): Here we use default CUDA stream to set a Python numpy
-  // array to a GPU Tensor. It's better to manage CDUA stream unifiedly.
-  paddle::platform::GpuMemcpySync(dst, array.data(), sizeof(T) * array.size(),
-                                  cudaMemcpyHostToDevice);
+
+  platform::DeviceContextPool &pool = platform::DeviceContextPool::Get();
+  auto dev_ctx =
+      static_cast<const platform::CUDADeviceContext *>(pool.Borrow(place));
+  paddle::platform::GpuMemcpyAsync(dst, array.data(), sizeof(T) * array.size(),
+                                   cudaMemcpyHostToDevice, dev_ctx->stream());
 }
 #endif
 

paddle/scripts/CMakeLists.txt

@@ -5,11 +5,3 @@ configure_file(submit_local.sh.in
 install(FILES ${CMAKE_CURRENT_BINARY_DIR}/paddle DESTINATION bin
         PERMISSIONS OWNER_EXECUTE OWNER_WRITE OWNER_READ
             GROUP_EXECUTE GROUP_READ WORLD_EXECUTE WORLD_READ)
-
-configure_file(tools/usage_stat/usage.sh
-    paddle_usage
-    @ONLY)
-
-install(FILES ${CMAKE_CURRENT_BINARY_DIR}/paddle_usage DESTINATION opt/paddle/bin
-        PERMISSIONS OWNER_EXECUTE OWNER_WRITE OWNER_READ
-            GROUP_EXECUTE GROUP_READ WORLD_EXECUTE WORLD_READ)

paddle/scripts/submit_local.sh.in

@@ -165,9 +165,6 @@ case "$1" in
     "make_diagram")
         python -m paddle.utils.make_model_diagram ${@:2}
         ;;
-    "usage")
-        $PADDLE_BIN_PATH/paddle_usage ${@:2}
-        ;;
     "version")
         version
         ;;

paddle/scripts/tools/usage_stat/usage.sh

-#!/bin/bash
-
-ARGPARSE=`getopt -o u:vin:l:e: --long git-user:,help,dry-run,task-name:,log-file:,exit-code:  -- "$@"`
-KEEP_ANONYMOUS="A_USER_DOES_NOT_TELL_US"
-# paddle config home dir, same as paddle
-PADDLE_CONF_HOME="$HOME/.config/paddle"
-# api url, mirror url(s) will be append later
-PD_URLS="http://api.paddlepaddle.org/version"
-
-usage()
-{
-    echo "Usage: `basename $0` [options]"
-    echo "Options:"
-    echo "  -e, --exit-code=EXIT_CODE         The train/predict process's exit code"
-    echo "  -l, --log-file=LOG_FILE_PATH      Read which log file to get the duration of process"
-    echo "  -n, --task-name=TASK_NAME         The name of demo or example"
-    echo "  -u, --git-user=GITHUB_USER        provide contact info, like username or email"
-    echo "  -v, -i                            Verbose output and interact with user when necessary"
-    echo " --help                             display this help message"
-}
-
-eval set -- "${ARGPARSE}"
-while true; do
-    case "$1" in
-        -l|--log-file)
-            log_file=$2
-            shift 2
-            ;;
-        -e|--exit-code)
-            exit_code=$2
-            shift 2
-            ;;
-        -u|--git-user)
-            github_user=$2
-            shift 2
-            ;;
-        -n|--task-name)
-            task=$2
-            shift 2
-            ;;
-        -v|-i)
-            v=1
-            shift
-            ;;
-        --dry-run)
-            dry_run=1
-            shift
-            ;;
-        --)
-            shift
-            break
-            ;;
-        --help)
-            usage
-            exit 0
-            ;;
-        *)
-            echo "Invalid option $1"
-            usage
-            exit 1
-            ;;
-    esac
-done
-
-# parse the log_file to get the time costs
-if [ -s "${log_file}" ]; then
-    duration=`awk 'BEGIN{day=0;last_sec=0;min_sec=0;max_sec=0;}
-    {if(index($2,":")==3){
-        t=substr($2,1,8);
-        sec=day*86400+substr(t,1,2)*3600+substr(t,4,2)*60+substr(t,7,2);
-        if(sec<last_sec-600){day+=1;sec+=86400;}
-        last_sec=sec;
-        if(min_sec==0 || min_sec>sec){min_sec=sec;}
-        if(max_sec==0 || max_sec<sec){max_sec=sec;}
-    }}
-    END{print max_sec-min_sec}' ${log_file}`
-else
-    duration=-1
-fi
-if [ "${v}" = "1" ]; then echo "duration: ${duration}"; fi
-
-# try find the user/email if not given
-if [ -z "${github_user}" ]; then
-    # search for cached username
-    if [ -s "${PADDLE_CONF_HOME}/github_user" ]; then
-        if [ "${v}" = "1" ]; then echo "read github_user from cache..."; fi
-        github_user=`cat ${PADDLE_CONF_HOME}/github_user`
-    else
-        # search the github-user from git config
-        if [ "${v}" = "1" ]; then echo "read github_user from git..."; fi
-        git_username=`git config --get user.name 2>/dev/null`
-        git_url=`git config --get remote.origin.url 2>/dev/null`
-        if [ "`echo ${git_url} | cut -b 1-19`" = "https://github.com/" ]; then
-            # under a git url, like https://github.com/user_xxx/proj_yyy.git
-            if [ "${v}" = "1" ]; then echo " from github url..."; fi
-            github_user=`echo ${git_url} | cut -d "/" -f 4`
-            if [ "${github_user}" = "PaddlePaddle" ]; then
-                github_user=
-            fi
-        fi
-        if [ -n "${git_username}" -a -z "${github_user}" ]; then
-            if [ "${v}" = "1" ]; then echo " from global git username..."; fi
-            github_user=${git_username}
-        fi
-    fi
-fi
-# allow user to set the user name, if it's not found
-if [ -z "${github_user}" -a "${v}" = "1" ]; then
-    read -p "Please input your github username or email, or just return to keep this feedback anonymous:"
-    github_user=${REPLY}
-    if [ -z "${github_user}" ]; then
-        # empty input, consider as one anonymous user
-        github_user="${KEEP_ANONYMOUS}"
-    fi
-fi
-if [ -n "${github_user}" -a -z "${dry_run}" ]; then
-    # valid user and not in dry-run mode, then save to cache
-    mkdir -p ${PADDLE_CONF_HOME}
-    echo "${github_user}" >${PADDLE_CONF_HOME}/github_user
-fi
-if [ "${v}" = "1" ]; then echo "username: ${github_user}"; fi
-if [ "${github_user}" = "${KEEP_ANONYMOUS}" ]; then
-    # anonymous user should keep the var empty.
-    github_user=
-fi
-
-# read local paddle version
-paddle_version=`paddle version | grep PaddlePaddle | head -n1 | cut -d " " -f 2 | cut -d "," -f 1`
-if [ "${v}" = "1" ]; then echo "version:${paddle_version}"; fi
-
-# read local system time
-system_time=`date "+%Y%m%d%H%M%S"`
-if [ "${v}" = "1" ]; then echo "system time:${system_time}"; fi
-
-# make empty job_name as default value.
-if [ -z "${task}" ]; then
-    task="(unknown_task)"
-fi
-if [ "${v}" = "1" ]; then echo "task: ${task}"; fi
-
-# concat the curl command
-params="content={\"data_type\":\"usage\",\
-\"system_time\":${system_time},\"paddle_version\":\"${paddle_version}\",\
-\"github_user\":\"${github_user}\",\"job_name\":\"${task}\",\
-\"duration\":${duration},\"exit_code\":\"${exit_code}\"\
-}&type=1"
-curl_cmd_prefix="curl -m 5 -X POST -d ${params}\
- -b ${PADDLE_CONF_HOME}/paddle.cookie -c ${PADDLE_CONF_HOME}/paddle.cookie "
-
-if [ "${dry_run}" = "1" ]; then
-    first_url=`echo ${PD_URLS} | cut -d " " -f 1`
-    echo "(dry-run mode)curl command: ${curl_cmd_prefix} ${first_url}"
-    exit 0
-else
-    for u in ${PD_URLS}; do
-        curl_cmd="${curl_cmd_prefix} ${u}"
-        if [ "${v}" = "1" ]; then echo "run: ${curl_cmd}"; fi
-        ${curl_cmd} >/dev/null 2>&1
-        if [ $? -eq 0 ]; then
-            if [ "${v}" = "1" ]; then echo "upload OK!"; fi
-            exit 0
-        else
-            if [ "${v}" = "1" ]; then echo "upload failed...try next"; fi
-        fi
-    done
-    if [ "${v}" = "1" ]; then echo "all urls tried but all failed...exit"; fi
-    exit 1
-fi

paddle/testing/CMakeLists.txt

@@ -6,7 +6,6 @@ if(WITH_TESTING)
   add_library(paddle_test_util STATIC TestUtil.cpp)
   add_dependencies(paddle_test_util paddle_proto ${external_project_dependencies})
   if(NOT MOBILE_INFERENCE)
-    add_library(paddle_gtest_main STATIC paddle_gtest_main.cc)
-    add_dependencies(paddle_gtest_main paddle_memory gtest gflags)
+    cc_library(paddle_gtest_main SRCS paddle_gtest_main.cc DEPS init paddle_memory gtest gflags)
   endif()
 endif()

paddle/testing/paddle_gtest_main.cc

@@ -13,8 +13,10 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include <cstring>
+
 #include "gflags/gflags.h"
 #include "gtest/gtest.h"
+#include "paddle/framework/init.h"
 #include "paddle/memory/memory.h"
 
 int main(int argc, char** argv) {
@@ -32,8 +34,11 @@ int main(int argc, char** argv) {
   google::ParseCommandLineFlags(&new_argc, &new_argv_address, false);
   testing::InitGoogleTest(&argc, argv);
   paddle::memory::Used(paddle::platform::CPUPlace());
+  std::vector<std::string> devs = {"CPU"};
 #ifdef PADDLE_WITH_CUDA
-  paddle::memory::Used(paddle::platform::GPUPlace(0));
+  paddle::memory::Used(paddle::platform::CUDAPlace(0));
+  devs.push_back("GPU:0");
 #endif
+  paddle::framework::InitDevices(devs);
   return RUN_ALL_TESTS();
 }

python/paddle/trainer_config_helpers/layers.py

@@ -270,7 +270,7 @@ class LayerType(object):
     @staticmethod
     def is_layer_type(type_name):
         """
-        If type_name is a layer type.
+        Whether type_name is a layer type.
 
         :param type_name: layer type name. Because layer type enumerations are
                           strings.
@@ -441,7 +441,7 @@ def full_matrix_projection(input, size=0, param_attr=None):
        with mixed_layer(size=100) as m:
            m += full_matrix_projection(input=layer)
 
-    2. When used as an independant object like this, you must set the size:
+    2. When used as an independent object like this, you must set the size:
 
     .. code-block:: python
 
@@ -451,11 +451,11 @@ def full_matrix_projection(input, size=0, param_attr=None):
 
     :param input: The input of this layer.
     :type input: LayerOutput
-    :param size: The parameter size. Means the width of parameter.
+    :param size: The dimension of this layer.
     :type size: int
-    :param param_attr: Parameter config, None if use default.
+    :param param_attr: The parameter attribute. See ParameterAttribute for details.
     :type param_attr: ParameterAttribute
-    :return: A FullMatrixProjection Object.
+    :return: FullMatrixProjection Object.
     :rtype: FullMatrixProjection
     """
     proj = FullMatrixProjection(
@@ -468,12 +468,12 @@ def full_matrix_projection(input, size=0, param_attr=None):
 def trans_full_matrix_projection(input, size=0, param_attr=None):
     """
     Different from full_matrix_projection, this projection performs matrix
-    multiplication, using transpose of weight.
+    multiplication, using the transpose of weight.
 
     ..  math::
         out.row[i] += in.row[i] * w^\mathrm{T}
 
-    :math:`w^\mathrm{T}` means transpose of weight.
+    :math:`w^\mathrm{T}` means the transpose of weight.
     The simply usage is:
 
     .. code-block:: python
@@ -489,9 +489,9 @@ def trans_full_matrix_projection(input, size=0, param_attr=None):
     :type input: LayerOutput
     :param size: The parameter size. Means the width of parameter.
     :type size: int
-    :param param_attr: Parameter config, None if use default.
+    :param param_attr: The parameter attribute. See ParameterAttribute for details.
     :type param_attr: ParameterAttribute
-    :return: A TransposedFullMatrixProjection Object.
+    :return: TransposedFullMatrixProjection Object.
     :rtype: TransposedFullMatrixProjection
     """
     proj = TransposedFullMatrixProjection(
@@ -521,7 +521,7 @@ def table_projection(input, size=0, param_attr=None):
        with mixed_layer(size=100) as m:
            m += table_projection(input=layer)
 
-    2. When used as an independant object like this, you must set the size:
+    2. When used as an independent object like this, you must set the size:
 
     .. code-block:: python
 
@@ -532,11 +532,11 @@ def table_projection(input, size=0, param_attr=None):
 
     :param input: The input of this layer, which must contains id fields.
     :type input: LayerOutput
-    :param size: The parameter size. Means the width of parameter.
+    :param size: The dimension of the output.
     :type size: int
-    :param param_attr: Parameter config, None if use default.
+    :param param_attr: The parameter attribute. See ParameterAttribute for details.
     :type param_attr: ParameterAttribute
-    :return: A TableProjection Object.
+    :return: TableProjection Object.
     :rtype: TableProjection
     """
     proj = TableProjection(
@@ -547,7 +547,7 @@ def table_projection(input, size=0, param_attr=None):
 
 def identity_projection(input, offset=None, size=None):
     """
-    1. IdentityProjection if offset=None. It performs:
+    1. If offset=None, it performs IdentityProjection as follows:
 
     .. math::
        out.row[i] += in.row[i]
@@ -559,9 +559,8 @@ def identity_projection(input, offset=None, size=None):
        proj = identity_projection(input=layer)
 
 
-    2. IdentityOffsetProjection if offset!=None. It likes IdentityProjection,
-    but layer size may be smaller than input size.
-    It select dimesions [offset, offset+layer_size) from input:
+    2. If offset!=None, It executes IdentityOffsetProjection and takes the
+       elements of the input in the range [offset, offset+size) as output.
 
     .. math::
        out.row[i] += in.row[i + \\textrm{offset}]
@@ -573,14 +572,20 @@ def identity_projection(input, offset=None, size=None):
        proj = identity_projection(input=layer,
                                   offset=10)
 
-    Note that both of two projections should not have any parameter.
+    Note that neither of the projections have trainable parameter.
 
     :param input: The input of this layer.
     :type input: LayerOutput
-    :param offset: Offset, None if use default.
+    :param offset: The offset from the start of the input. The input's
+                   elements in the range [offset, offset+size) will be
+                   taken as output. If this parameter is not set or set
+                   to None, the output will be the same as the input.
     :type offset: int
-    :return: A IdentityProjection or IdentityOffsetProjection object
-    :rtype: IdentityProjection or IdentityOffsetProjection
+    :param size: The dimension of this layer. It will be neglected
+                 when offset is None or not set.
+    :type size: int
+    :return: IdentityProjection or IdentityOffsetProjection object
+    :rtype: IdentityProjection | IdentityOffsetProjection
     """
     if offset is None:
         proj = IdentityProjection(input_layer_name=input.name)
@@ -596,8 +601,8 @@ def identity_projection(input, offset=None, size=None):
 
 def slice_projection(input, slices):
     """
-    slice_projection can slice the input value into multiple parts,
-    and then select some of them to merge into a new output.
+    slice_projection slices the input value into multiple parts,
+    then selects and merges some of them into a new output.
 
     .. math::
        output = [input.slices()]
@@ -608,15 +613,13 @@ def slice_projection(input, slices):
 
        proj = slice_projection(input=layer, slices=[(0, 10), (20, 30)])
 
-    Note that slice_projection should not have any parameter.
+    Note that slice_projection has no trainable parameter.
 
     :param input: The input of this layer.
     :type input: LayerOutput
-    :param slices: An array of slice parameters.
-                   Each slice contains the start and end offsets based
-                   on the input.
-    :type slices: pair of int
-    :return: A SliceProjection object
+    :param slices: A list of start and end offsets of each slice.
+    :type slices: list of tuple
+    :return: SliceProjection object.
     :rtype: SliceProjection
     """
     assert len(slices) >= 1
@@ -636,8 +639,7 @@ def slice_projection(input, slices):
 @wrap_param_attr_default()
 def scaling_projection(input, param_attr=None):
     """
-    scaling_projection multiplies the input with a scalar parameter and add to
-    the output.
+    scaling_projection multiplies the input with a scalar parameter.
 
     .. math::
        out += w * in
@@ -650,9 +652,9 @@ def scaling_projection(input, param_attr=None):
 
     :param input: The input of this layer.
     :type input: LayerOutput
-    :param param_attr: Parameter config, None if use default.
+    :param param_attr: The parameter attribute. See ParameterAttribute for details.
     :type param_attr: ParameterAttribute
-    :return: A ScalingProjection object
+    :return: ScalingProjection object.
     :rtype: ScalingProjection
     """
     proj = ScalingProjection(input_layer_name=input.name, **param_attr.attr)
@@ -663,8 +665,8 @@ def scaling_projection(input, param_attr=None):
 @wrap_param_attr_default()
 def dotmul_projection(input, param_attr=None):
     """
-    DotMulProjection with a layer as input.
-    It performs element-wise multiplication with weight.
+    DotMulProjection takes a layer as input and performs
+    element-wise multiplication with weight.
 
     ..  math::
         out.row[i] += in.row[i] .* weight
@@ -679,9 +681,9 @@ def dotmul_projection(input, param_attr=None):
 
     :param input: The input of this layer.
     :type input: LayerOutput
-    :param param_attr: Parameter config, None if use default.
+    :param param_attr: The parameter attribute. See ParameterAttribute for details.
     :type param_attr: ParameterAttribute
-    :return: A DotMulProjection Object.
+    :return: DotMulProjection object.
     :rtype: DotMulProjection
     """
     proj = DotMulProjection(
@@ -698,7 +700,7 @@ def dotmul_operator(a=None, b=None, scale=1, **kwargs):
        out.row[i] += scale * (a.row[i] .* b.row[i])
 
     where :math:`.*` means element-wise multiplication, and
-    scale is a config scalar, its default value is one.
+    scale is a config scalar, its default value is 1.
 
     The example usage is:
 
@@ -706,13 +708,13 @@ def dotmul_operator(a=None, b=None, scale=1, **kwargs):
 
        op = dotmul_operator(a=layer1, b=layer2, scale=0.5)
 
-    :param a: Input layer1
+    :param a: The first input of this layer.
     :type a: LayerOutput
-    :param b: Input layer2
+    :param b: The second input of this layer.
     :type b: LayerOutput
-    :param scale: config scalar, default value is one.
+    :param scale: A scalar to scale the product. Its default value is 1.
     :type scale: float
-    :return: A DotMulOperator Object.
+    :return: DotMulOperator object.
     :rtype: DotMulOperator
     """
     if 'x' in kwargs or 'y' in kwargs:
@@ -738,28 +740,29 @@ def context_projection(input,
     """
     Context Projection.
 
-    It just simply reorganizes input sequence, combines "context_len" sequence
-    to one context from context_start. "context_start" will be set to
-    -(context_len - 1) / 2 by default. If context position out of sequence
+    It just reorganizes input sequence, combines "context_len" elements of the
+    sequence to one context from context_start. "context_start" will be set to
+    -(context_len - 1) / 2 by default. When context position is out of sequence
     length, padding will be filled as zero if padding_attr = False, otherwise
     it is trainable.
 
-    For example, origin sequence is [A B C D E F G], context len is 3, then
-    after context projection and not set padding_attr, sequence will
+    For example, origin sequence is [A B C D E F G], context len is 3, padding_attr
+    is not set, then after context projection, sequence will
     be [ 0AB ABC BCD CDE DEF EFG FG0 ].
 
     :param input: The input of this layer, which should be a sequence.
     :type input: LayerOutput
-    :param context_len: context length.
+    :param context_len: The length of the context.
     :type context_len: int
-    :param context_start: context start position. Default is
+    :param context_start: The start position of the context. The default value is
                           -(context_len - 1)/2
     :type context_start: int
-    :param padding_attr: Padding Parameter Attribute. If false, it means padding
-                         always be zero. Otherwise Padding is learnable, and
-                         parameter attribute is set by this parameter.
+    :param padding_attr: Parameter attribute of the padding. If the parameter is
+                         set to False, padding will be zero. In other cases, the
+                         padding is trainable, and its parameter attribute is set
+                         by this parameter.
     :type padding_attr: bool | ParameterAttribute
-    :return: Projection
+    :return: Projection object.
     :rtype: Projection
     """
     context_start = -(
@@ -791,10 +794,9 @@ class MixedLayerType(LayerOutput):
 
     def __init__(self, name, size, act, bias_attr, layer_attr, parents=None):
         """
-        Ctor.
-        :param name: layer name.
+        :param name: The name of this layer.
         :type name: basestring
-        :param size: layer size.
+        :param size: The dimension of this layer.
         :type size: int
         :param act: Activation type.
         :type act: BaseActivation
@@ -802,8 +804,9 @@ class MixedLayerType(LayerOutput):
                           whose type is not ParameterAttribute, no bias is defined. If the
                           parameter is set to True, the bias is initialized to zero.
         :type bias_attr: ParameterAttribute | None | bool | Any
-        :param layer_attr: Extra Layer Attribute.
-        :type layer_attr: ExtraLayerAttribute or None
+        :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for
+                           details.
+        :type layer_attr: ExtraLayerAttribute | None
         """
         LayerOutput.__init__(
             self,
@@ -868,12 +871,12 @@ def mixed_layer(size=0,
                 bias_attr=False,
                 layer_attr=None):
     """
-    Mixed Layer. A mixed layer will add all inputs together, then activate.
-    Each inputs is a projection or operator.
+    Mixed Layer. A mixed layer will add all inputs together, then activate the sum.
+    Each input is a projection or operator.
 
     There are two styles of usages.
 
-    1. When not set inputs parameter, use mixed_layer like this:
+    1. When the parameter input is not set, use mixed_layer like this:
 
     .. code-block:: python
 
@@ -889,21 +892,21 @@ def mixed_layer(size=0,
                        input=[full_matrix_projection(input=layer1),
                               full_matrix_projection(input=layer2)])
 
-    :param name: mixed layer name. Can be referenced by other layer.
+    :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param size: layer size.
+    :param size: The dimension of this layer.
     :type size: int
-    :param input: The input of this layer. It is an optional parameter. If set,
-                  then this function will just return layer's name.
+    :param input: The input of this layer. It is an optional parameter.
     :param act: Activation Type. LinearActivation is the default activation.
     :type act: BaseActivation
     :param bias_attr: The bias attribute. If the parameter is set to False or an object
                       whose type is not ParameterAttribute, no bias is defined. If the
                       parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
-    :param layer_attr: The extra layer config. Default is None.
+    :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for
+                       details.
     :type layer_attr: ExtraLayerAttribute
-    :return: MixedLayerType object can add inputs or layer name.
+    :return: MixedLayerType object.
     :rtype: MixedLayerType
     """
 
@@ -938,14 +941,15 @@ def data_layer(name, size, depth=None, height=None, width=None,
 
     :param name: The name of this layer.
     :type name: basestring
-    :param size: Size of this data layer.
+    :param size: The dimension of this data layer.
     :type size: int
-    :param height: Height of this data layer, used for image
+    :param height: The height of the input image data.
     :type height: int | None
-    :param width: Width of this data layer, used for image
+    :param width: The width of the input image data.
     :type width: int | None
-    :param layer_attr: Extra Layer Attribute.
-    :type layer_attr: ExtraLayerAttribute.
+    :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for
+                       details.
+    :type layer_attr: ExtraLayerAttribute
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -978,14 +982,15 @@ def embedding_layer(input, size, name=None, param_attr=None, layer_attr=None):
 
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param input: The input of this layer, which must be Index Data.
+    :param input: The input of this layer, whose type must be Index Data.
     :type input: LayerOutput
-    :param size: The embedding dimension.
+    :param size: The dimension of the embedding vector.
     :type size: int
     :param param_attr: The embedding parameter attribute. See ParameterAttribute
                       for details.
-    :type param_attr: ParameterAttribute | None
-    :param layer_attr: Extra layer Config. Default is None.
+    :type param_attr: ParameterAttribute
+    :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for
+                       details.
     :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -1013,7 +1018,7 @@ def fc_layer(input,
              bias_attr=None,
              layer_attr=None):
     """
-    Helper for declare fully connected layer.
+    The fully connected layer.
 
     The example usage is:
 
@@ -1035,17 +1040,18 @@ def fc_layer(input,
     :type name: basestring
     :param input: The input of this layer.
     :type input: LayerOutput | list | tuple
-    :param size: The layer dimension.
+    :param size: The dimension of this layer.
     :type size: int
     :param act: Activation Type. TanhActivation is the default activation.
     :type act: BaseActivation
-    :param param_attr: The Parameter Attribute|list.
+    :param param_attr: The parameter attribute. See ParameterAttribute for details.
     :type param_attr: ParameterAttribute
     :param bias_attr: The bias attribute. If the parameter is set to False or an object
                       whose type is not ParameterAttribute, no bias is defined. If the
                       parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
-    :param layer_attr: Extra Layer config.
+    :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for
+                       details.
     :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -1086,13 +1092,15 @@ def fc_layer(input,
 @wrap_name_default("print")
 def printer_layer(input, format=None, name=None):
     """
-    Print the output value of input layers. This layer is useful for debugging.
+    Print the output value of the layers specified by the parameter input.
+    This layer is useful for debugging.
 
     :param name: The name of this layer. It is optional.
     :type name: basestring
     :param input: The input of this layer.
     :type input: LayerOutput | list | tuple
-    :return: LayerOutput
+    :return: LayerOutput object.
+    :rtype: LayerOutput
     """
     if isinstance(input, LayerOutput):
         input = [input]
@@ -1135,11 +1143,12 @@ def priorbox_layer(input,
     :param aspect_ratio: The aspect ratio.
     :type aspect_ratio: list
     :param variance: The bounding box variance.
-    :type min_size: The min size of the priorbox width/height.
+    :type min_size: The minimum size of the priorbox width/height.
     :param min_size: list
-    :type max_size: The max size of the priorbox width/height. Could be NULL.
+    :type max_size: The maximum size of the priorbox width/height. It could be NULL.
     :param max_size: list
-    :return: LayerOutput
+    :return: LayerOutput object.
+    :rtype: LayerOutput
     """
     # plus one for ratio 1.
     num_filters = (len(aspect_ratio) * 2 + 1 + len(max_size)) * 4
@@ -1177,7 +1186,7 @@ def multibox_loss_layer(input_loc,
 
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param input_loc: The input predict locations.
+    :param input_loc: The input predicted locations.
     :type input_loc: LayerOutput | List of LayerOutput
     :param input_conf: The input priorbox confidence.
     :type input_conf: LayerOutput | List of LayerOutput
@@ -1189,13 +1198,15 @@ def multibox_loss_layer(input_loc,
     :type num_classes: int
     :param overlap_threshold: The threshold of the overlap.
     :type overlap_threshold: float
-    :param neg_pos_ratio: The ratio of the negative bbox to the positive bbox.
+    :param neg_pos_ratio: The ratio of the negative bounding box to
+                          the positive bounding box.
     :type neg_pos_ratio: float
-    :param neg_overlap: The negative bbox overlap threshold.
+    :param neg_overlap: The negative bounding box overlap threshold.
     :type neg_overlap: float
     :param background_id: The background class index.
     :type background_id: int
-    :return: LayerOutput
+    :return: LayerOutput object.
+    :rtype: LayerOutput
     """
     if isinstance(input_loc, LayerOutput):
         input_loc = [input_loc]
@@ -1258,19 +1269,20 @@ def detection_output_layer(input_loc,
     :type input_conf: LayerOutput | List of LayerOutput.
     :param priorbox: The input priorbox location and the variance.
     :type priorbox: LayerOutput
-    :param num_classes: The number of the classification.
+    :param num_classes: The number of the classes.
     :type num_classes: int
     :param nms_threshold: The Non-maximum suppression threshold.
     :type nms_threshold: float
-    :param nms_top_k: The bbox number kept of the NMS's output
+    :param nms_top_k: The bounding boxes number kept of the NMS's output.
     :type nms_top_k: int
-    :param keep_top_k: The bbox number kept of the layer's output
+    :param keep_top_k: The bounding boxes number kept of the layer's output.
     :type keep_top_k: int
-    :param confidence_threshold: The classification confidence threshold
+    :param confidence_threshold: The classification confidence threshold.
     :type confidence_threshold: float
     :param background_id: The background class index.
     :type background_id: int
-    :return: LayerOutput
+    :return: LayerOutput object.
+    :rtype: LayerOutput
     """
     if isinstance(input_loc, LayerOutput):
         input_loc = [input_loc]
@@ -1326,7 +1338,7 @@ def roi_pool_layer(input,
     A layer used by Fast R-CNN to extract feature maps of ROIs from the last
     feature map.
 
-    :param name: The Layer Name.
+    :param name: The name of this layer. It is optional.
     :type name: basestring
     :param input: The input layer.
     :type input: LayerOutput.
@@ -1338,9 +1350,10 @@ def roi_pool_layer(input,
     :type pooled_height: int
     :param spatial_scale: The spatial scale between the image and feature map.
     :type spatial_scale: float
-    :param num_channels: number of input channel.
+    :param num_channels: The number of the input channels.
     :type num_channels: int
-    :return: LayerOutput
+    :return: LayerOutput object.
+    :rtype: LayerOutput
     """
     if num_channels is None:
         assert input.num_filters is not None
@@ -1361,18 +1374,19 @@ def roi_pool_layer(input,
 @wrap_name_default("cross_channel_norm")
 def cross_channel_norm_layer(input, name=None, param_attr=None):
     """
-    Normalize a layer's output. This layer is necessary for ssd.
-    This layer applys normalize across the channels of each sample to
-    a conv layer's output and scale the output by a group of trainable
-    factors which dimensions equal to the channel's number.
+    Normalize a layer's output. This layer is necessary for ssd. This
+    layer applys normalization across the channels of each sample to
+    a convolutional layer's output and scales the output by a group of
+    trainable factors whose dimensions equal to the channel's number.
 
     :param name: The name of this layer. It is optional.
     :type name: basestring
     :param input: The input of this layer.
     :type input: LayerOutput
-    :param param_attr: The Parameter Attribute|list.
+    :param param_attr: The parameter attribute. See ParameterAttribute for details.
     :type param_attr: ParameterAttribute
-    :return: LayerOutput
+    :return: LayerOutput object.
+    :rtype: LayerOutput
     """
     assert input.num_filters is not None
     Layer(
@@ -1413,12 +1427,9 @@ def pooling_layer(input,
     Pooling layer for sequence inputs, not used for Image.
 
     If stride > 0, this layer slides a window whose size is determined by stride,
-    and return the pooling value of the window as the output. Thus, a long sequence
-    will be shorten.
-
-    The parameter stride specifies the intervals at which to apply the pooling
-    operation. Note that for sequence with sub-sequence, the default value
-    of stride is -1.
+    and returns the pooling value of the sequence in the window as the output. Thus,
+    a long sequence will be shortened. Note that for sequence with sub-sequence, the
+    default value of stride is -1.
 
     The example usage is:
 
@@ -1435,16 +1446,16 @@ def pooling_layer(input,
     :type name: basestring
     :param input: The input of this layer.
     :type input: LayerOutput
-    :param pooling_type: Type of pooling, MaxPooling(default), AvgPooling,
-                         SumPooling, SquareRootNPooling.
+    :param pooling_type: Type of pooling. MaxPooling is the default pooling.
     :type pooling_type: BasePoolingType | None
     :param stride: The step size between successive pooling regions.
-    :type stride: Int
+    :type stride: int
     :param bias_attr: The bias attribute. If the parameter is set to False or an object
                       whose type is not ParameterAttribute, no bias is defined. If the
                       parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
-    :param layer_attr: The Extra Attributes for layer, such as dropout.
+    :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for
+                       details.
     :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -6618,7 +6629,7 @@ def row_conv_layer(input,
     .. math::
 
         r_{t,r} = \sum_{j=1}^{k + 1} {w_{i,j}h_{t+j-1, i}}
-                  \quad \text{for} \quad  (1 \leq i \leq d)
+                  \quad \\text{for} \quad  (1 \leq i \leq d)
 
     Note:
         The `context_len` is `k + 1`. That is to say, the lookahead step
@@ -6767,7 +6778,7 @@ def gated_unit_layer(input,
     The gated unit layer implements a simple gating mechanism over the input.
     The input :math:`X` is first projected into a new space :math:`X'`, and
     it is also used to produce a gate weight :math:`\sigma`. Element-wise
-    product between :match:`X'` and :math:`\sigma` is finally returned.
+    product between :math:`X'` and :math:`\sigma` is finally returned.
 
     Reference:
         `Language Modeling with Gated Convolutional Networks
@@ -7463,7 +7474,7 @@ def factorization_machine(input,
     Factorization Machine with the formula:
 
     .. math::
-        y = \sum_{i=1}^{n-1}\sum_{j=i+1}^n\langle v_i, v_j \rangle x_i x_j
+        y = \sum_{i=1}^{n-1}\sum_{j=i+1}^n\langle v_i, v_j \\rangle x_i x_j
 
     Note:
         X is the input vector with size n. V is the factor matrix. Each row of V

python/paddle/v2/fluid/__init__.py

@@ -15,14 +15,15 @@ import backward
 import regularizer
 from param_attr import ParamAttr
 from data_feeder import DataFeeder
-from core import LoDTensor, CPUPlace, GPUPlace
+from core import LoDTensor, CPUPlace, CUDAPlace
+from distribute_transpiler import DistributeTranspiler
 import clip
 
 Tensor = LoDTensor
 __all__ = framework.__all__ + executor.__all__ + [
     'io', 'initializer', 'layers', 'nets', 'optimizer', 'backward',
-    'regularizer', 'LoDTensor', 'CPUPlace', 'GPUPlace', 'Tensor', 'ParamAttr'
-    'DataFeeder', 'clip'
+    'regularizer', 'LoDTensor', 'CPUPlace', 'CUDAPlace', 'Tensor', 'ParamAttr'
+    'DataFeeder', 'clip', 'DistributeTranspiler'
 ]
 
 
@@ -41,5 +42,10 @@ def __read_gflags_from_env__():
     core.init_gflags([sys.argv[0]] +
                      ["--tryfromenv=" + ",".join(read_env_flags)])
 
+    if core.is_compile_gpu():
+        core.init_devices(["CPU", "GPU:0"])
+    else:
+        core.init_devices(["CPU"])
+
 
 __read_gflags_from_env__()

python/paddle/v2/fluid/distribute_transpiler.py

+import framework
+from framework import Program, default_main_program, Parameter, Variable
+import optimizer
+from layer_helper import LayerHelper
+
+
+def hash_name_to_server(params_grads, pserver_endpoints):
+    """
+    :param param_grads:
+    :return: a map of pserver endpoint -> 
+                    params -> [param list]
+                    grads  -> [grad list]
+    """
+
+    def _hash_param(param_name, total):
+        return hash(param_name) % total
+
+    param_grad_map = dict()
+    for param, grad in params_grads:
+        if param.trainable is True and grad is not None:
+            server_id = _hash_param(param.name, len(pserver_endpoints))
+            server_for_param = pserver_endpoints[server_id]
+            if not param_grad_map.has_key(server_for_param):
+                param_grad_map[server_for_param] = {"params": [], "grads": []}
+            param_grad_map[server_for_param]["params"].append(param)
+            param_grad_map[server_for_param]["grads"].append(grad)
+
+    return param_grad_map
+
+
+def round_robin(params_grads, pserver_endpoints):
+    assert (len(params_grads) > len(pserver_endpoints))
+
+    param_grad_map = dict()
+    pserver_idx = 0
+    for param, grad in params_grads:
+        if param.trainable is True:
+            server_for_param = pserver_endpoints[pserver_idx]
+            if not param_grad_map.has_key(server_for_param):
+                param_grad_map[server_for_param] = {"params": [], "grads": []}
+
+            param_grad_map[server_for_param]["params"].append(param)
+            param_grad_map[server_for_param]["grads"].append(grad)
+
+            pserver_idx += 1
+            if pserver_idx >= len(pserver_endpoints):
+                pserver_idx = 0
+    return param_grad_map
+
+
+class DistributeTranspiler:
+    def transpile(self,
+                  optimize_ops,
+                  params_grads,
+                  program=None,
+                  pservers="127.0.0.1:6174",
+                  trainers=1,
+                  split_method=round_robin):
+        """
+            Transpile the program to a distributed data-parallelism programs.
+
+            The main_program will be transform to use a remote parameter server
+            to do parameter optimization. And the optimization graph will be put
+            in to a parameter server program.
+
+            Use different methods to split trainable varialbles to different
+            parameter servers.
+
+            Example to run:
+
+            exe = fluid.Executor(place)
+            t = fluid.DistributeTranspiler()
+            t.transpile(optimize_ops, params_grads, pservers="127.0.0.1:6174", trainers=1)
+
+            pserver_endpoint = os.getenv("PSERVER")
+            if pserver_endpoint:
+                pserver_prog = t.get_pserver_program(pserver_endpoint, optimize_ops)
+                exe.run(fluid.default_startup_program())
+                exe.run(pserver_prog)
+            else:
+                feeder = fluid.DataFeeder(feed_list=[images, label], place=place)
+                exe.run(fluid.default_startup_program())
+
+                for pass_id in range(PASS_NUM):
+                    ...
+
+            :param optimize_ops: op list of optimization, should be the
+                                 return value of Optimizer.minimize
+            :type optimize_ops: list
+            :param program: program to optimize, default default_main_program
+            :param pservers: parameter server endpoints like "m1:6174,m2:6174"
+            :type pservers: string
+
+            :return: return a list of programs
+        """
+        if program is None:
+            program = default_main_program()
+        self.trainers = trainers
+        self._optimize_distributed(
+            optimize_ops,
+            program,
+            params_grads,
+            pservers=pservers,
+            trainers=trainers,
+            split_method=split_method)
+
+    def _clone_param(self, block, v):
+        assert isinstance(v, Parameter)
+        new_p = Parameter(
+            block=block,
+            shape=v.shape,
+            dtype=v.dtype,
+            type=v.type,
+            lod_level=v.lod_level,
+            stop_gradient=v.stop_gradient,
+            trainable=v.trainable,
+            optimize_attr=v.optimize_attr,
+            regularizer=v.regularizer,
+            name=v.name)
+        block.vars[new_p.name] = new_p
+
+    def _clone_var(self, block, var):
+        assert isinstance(var, Variable)
+        return block.create_var(
+            name=var.name,
+            shape=var.shape,
+            dtype=var.dtype,
+            type=var.type,
+            lod_level=var.lod_level,
+            persistable=var.persistable)
+
+    def _optimize_distributed(self, optimize_ops, program, params_and_grads,
+                              **kwargs):
+        if kwargs.has_key("split_method"):
+            split_method = kwargs["split_method"]
+        else:
+            split_method = round_robin
+
+        assert (callable(split_method))
+        pserver_endpoints = kwargs["pservers"].split(",")
+        self.param_grad_map = split_method(params_and_grads, pserver_endpoints)
+
+        send_op_ordered_inputs = []
+        epmap = []
+        for ep, v in self.param_grad_map.iteritems():
+            send_op_ordered_inputs.extend(v["grads"])
+            for i in v["grads"]:
+                epmap.append(ep)
+        send_op = program.global_block().append_op(
+            type="send",
+            inputs={"X": send_op_ordered_inputs
+                    },  # inputs is a list of tensors to be send
+            outputs={},
+            attrs={"endpoints": pserver_endpoints,
+                   "epmap": epmap})
+
+    def get_trainer_program(optimize_ops, program):
+        # remove optimize ops and add a send op to main_program
+        program.global_block().delete_ops(optimize_ops)
+
+    def _create_var_for_trainers(self, block, var, trainers):
+        var_list = []
+        for i in xrange(trainers):
+            var_each = block.create_var(
+                name="%s.trainer_%d" % (var.name, i),
+                psersistable=var.persistable,
+                dtype=var.dtype,
+                shape=var.shape)
+            var_list.append(var_each)
+        return var_list
+
+    def get_pserver_program(self, endpoint, optimize_ops):
+        pserver_program = Program()
+        for v in self.param_grad_map[endpoint]["params"]:
+            self._clone_param(pserver_program.global_block(), v)
+
+        optimize_sub_program = Program()
+        grad_var_names = [
+            var.name for var in self.param_grad_map[endpoint]["grads"]
+        ]
+        for opt_op in optimize_ops:
+            for _, var in opt_op.inputs.iteritems():
+                # NOTE: append operators to merge gradients from multiple
+                # trainers. If trainers == 1, this is not needed.
+                if self.trainers > 1 and var.name in grad_var_names:
+                    vars2merge = self._create_var_for_trainers(
+                        optimize_sub_program.global_block(), var, self.trainers)
+                    merged_var = optimize_sub_program.global_block().create_var(
+                        name=var.name,
+                        persistable=var.persistable,
+                        dtype=var.dtype,
+                        shape=var.shape)
+                    optimize_sub_program.global_block().append_op(
+                        type="sum",
+                        inputs={"X": vars2merge},
+                        outputs={"Out": merged_var})
+                    optimize_sub_program.global_block().append_op(
+                        type="scale",
+                        inputs={"X": merged_var},
+                        outputs={"Out": merged_var},
+                        attrs={"scale": 1.0 / float(self.trainers)})
+                else:
+                    optimize_sub_program.global_block().create_var(
+                        name=var.name,
+                        persistable=var.persistable,
+                        dtype=var.dtype,
+                        shape=var.shape)
+
+            if opt_op.inputs.has_key("Grad"):
+                if opt_op.inputs["Grad"].name in grad_var_names:
+                    print "appending ", opt_op.type, opt_op.inputs
+                    optimize_sub_program.global_block().append_op(
+                        type=opt_op.type,
+                        inputs=opt_op.inputs,
+                        outputs=opt_op.outputs,
+                        attrs=opt_op.attrs)
+            else:
+                optimize_sub_program.global_block().append_op(
+                    type=opt_op.type,
+                    inputs=opt_op.inputs,
+                    outputs=opt_op.outputs,
+                    attrs=opt_op.attrs)
+        pserver_program.global_block().append_op(
+            type="recv",
+            inputs={"RX":
+                    self.param_grad_map[endpoint]["grads"]},  # grads to recv
+            outputs={},
+            attrs={
+                "OptimizeProgram": optimize_sub_program.desc,
+                "endpoint": endpoint,
+                "ParamList":
+                [p.name for p in self.param_grad_map[endpoint]["params"]],
+                "GradList":
+                [p.name for p in self.param_grad_map[endpoint]["grads"]],
+                "Trainers": self.trainers
+            })
+        pserver_program.sync_with_cpp()
+        return pserver_program

python/paddle/v2/fluid/executor.py

 import numpy as np
 from . import core
-from framework import Program, default_main_program
+from framework import Program, default_main_program, Parameter, Variable
 
 __all__ = ['Executor', 'g_scope']
 
@@ -47,13 +47,14 @@ class Executor(object):
             act_places.append(p)
 
         # TODO(dzhwinter) : consider that our fluid tests all written in 
-        # GPUPlace(gpu_id), this will be changed in next PR.
+        # CUDAPlace(gpu_id), this will be changed in the future
         if core.is_compile_gpu():
             core.init_devices(["CPU", "GPU:0"])
         else:
             core.init_devices(["CPU"])
 
-        self.executor = core.Executor(act_places)
+        # TODO(dzhwinter) : only use the first place
+        self.executor = core.Executor(act_places[0])
         self.places = places
 
     def aslodtensor(self, data):
@@ -148,7 +149,7 @@ class Executor(object):
                 outputs={'Out': [fetch_var]},
                 attrs={'col': i})
 
-        self.executor.run(program.desc, scope, 0, True)
+        self.executor.run(program.desc, scope, 0, True, True)
         outs = [
             core.get_fetch_variable(scope, fetch_var_name, i)
             for i in xrange(len(fetch_list))

python/paddle/v2/fluid/framework.py

@@ -17,6 +17,10 @@ TEMP_VAR_NAME = core.kTempVarName()
 GRAD_VAR_SUFFIX = core.kGradVarSuffix()
 ZERO_VAR_SUFFIX = core.kZeroVarSuffix()
 
+USE_CPU = core.kUseCPU()
+USE_CUDNN = core.kUseMKLDNN()
+USE_MKLDNN = core.kUseMKLDNN()
+
 
 def grad_var_name(var_name):
     """
@@ -359,6 +363,10 @@ class Operator(object):
         """
         self.block = block
         self.desc = desc
+        # for clone a new operator
+        self.inputs = inputs
+        self.outputs = outputs
+        self.attrs = attrs
         if len(self.desc.type()) != 0:
             return
         if type is None:
@@ -389,6 +397,9 @@ class Operator(object):
                             % (in_proto.name, len(in_args)))
                     in_arg_names = []
                     for arg in in_args:
+                        if isinstance(arg, basestring):
+                            in_arg_names.append(arg)
+                        else:
                             in_arg_names.append(arg.name)
                     self.desc.set_input(in_proto.name, in_arg_names)
                 else:
@@ -430,13 +441,18 @@ class Operator(object):
                     continue
                 if isinstance(attrs[attr_name], Block):
                     self.desc.set_block_attr(attr_name, attrs[attr_name].desc)
+                elif isinstance(attrs[attr_name], core.BlockDesc) or \
+                   isinstance(attrs[attr_name], core.ProgramDesc):
+                    self.desc.set_serialized_attr(
+                        attr_name, attrs[attr_name].serialize_to_string())
                 else:
                     self.desc.set_attr(attr_name, attrs[attr_name])
 
         self.desc.check_attrs()
         no_kernel_op_set = {
             'feed', 'fetch', 'save', 'load', 'recurrent',
-            'rnn_memory_helper_grad', 'conditional_block', 'while'
+            'rnn_memory_helper_grad', 'conditional_block', 'while', 'send',
+            'recv'
         }
         if type not in no_kernel_op_set:
             self.desc.infer_var_type(self.block.desc)
@@ -582,6 +598,7 @@ class Block(object):
         self.vars = dict()  # var_name --> var
         self.ops = collections.deque()  # operator list
         self.program = program
+        self.removed_vars = dict()
 
     def __str__(self):
         return self.to_string(True)
@@ -638,6 +655,16 @@ class Block(object):
         self.ops.append(op)
         return op
 
+    def delete_ops(self, ops):
+        # remove from cpp
+        # FIXME(typhoonzero): remove only the first occuracy.
+        try:
+            start = list(self.ops).index(ops[0])
+            end = list(self.ops).index(ops[-1])
+        except Exception, e:
+            raise e
+        self.desc.remove_op(start, end)
+
     def prepend_op(self, *args, **kwargs):
         op_desc = self.desc.prepend_op()
         op = Operator(self, op_desc, *args, **kwargs)

python/paddle/v2/fluid/layer_helper.py

@@ -194,3 +194,9 @@ class LayerHelper(object):
         else:
             # For integer and boolean types, initialize with all zeros
             return Constant()
+
+    def is_instance(self, param_name, cls):
+        param = self.kwargs.get(param_name, None)
+        if not isinstance(param, cls):
+            raise TypeError("The input {0} parameter of method {1} must be {2}",
+                            param_name, self.layer_type, cls.__name__)

python/paddle/v2/fluid/layers/control_flow.py

@@ -3,6 +3,7 @@ from ..framework import Program, Variable, Operator
 from .. import core
 from tensor import assign, fill_constant
 import contextlib
+from ..registry import autodoc
 
 __all__ = [
     'split_lod_tensor', 'merge_lod_tensor', 'BlockGuard', 'StaticRNNGuard',
@@ -10,7 +11,7 @@ __all__ = [
     'max_sequence_len', 'topk', 'lod_tensor_to_array', 'array_to_lod_tensor',
     'increment', 'array_write', 'create_array', 'less_than', 'array_read',
     'shrink_memory', 'array_length', 'IfElse', 'DynamicRNN', 'ConditionalBlock',
-    'StaticRNN'
+    'StaticRNN', 'reorder_lod_tensor_by_rank'
 ]
 
 
@@ -1082,3 +1083,18 @@ class DynamicRNN(object):
         if self.status != DynamicRNN.IN_RNN:
             raise ValueError("{0} can only be invoked inside rnn block.".format(
                 method))
+
+
+@autodoc
+def reorder_lod_tensor_by_rank(x, rank_table):
+    helper = LayerHelper('reorder_lod_tensor_by_rank', **locals())
+    helper.is_instance('x', Variable)
+    helper.is_instance('rank_table', Variable)
+
+    out = helper.create_tmp_variable(dtype=x.dtype)
+    helper.append_op(
+        type='reorder_lod_tensor_by_rank',
+        inputs={'X': [x],
+                'RankTable': [rank_table]},
+        outputs={'Out': [out]})
+    return out

python/paddle/v2/fluid/layers/nn.py

@@ -13,7 +13,8 @@ __all__ = [
     'crf_decoding', 'cos_sim', 'cross_entropy', 'square_error_cost', 'accuracy',
     'chunk_eval', 'sequence_conv', 'conv2d', 'sequence_pool', 'pool2d',
     'batch_norm', 'beam_search_decode', 'conv2d_transpose', 'sequence_expand',
-    'lstm_unit', 'reduce_sum'
+    'lstm_unit', 'reduce_sum', 'reduce_mean', 'sequence_first_step',
+    'sequence_last_step'
 ]
 
 
@@ -41,7 +42,7 @@ def fc(input,
 
     .. math::
 
-        Out = Act\left({\sum_{i=0}^{N-1}W_iX_i + b}\right)
+        Out = Act({\sum_{i=0}^{N-1}W_iX_i + b})
 
     In the above equation:
 
@@ -162,8 +163,9 @@ def embedding(input, size, is_sparse=False, param_attr=None, dtype='float32'):
     Examples:
         .. code-block:: python
 
+          dict_size = len(dataset.ids)
           data = fluid.layers.data(name='ids', shape=[32, 32], dtype='float32')
-          fc = fluid.layers.embedding(input=data, size=16)
+          fc = fluid.layers.embedding(input=data, size=[dict_size, 16])
     """
 
     helper = LayerHelper('embedding', **locals())
@@ -575,8 +577,52 @@ def conv2d(input,
 def sequence_pool(input, pool_type, **kwargs):
     """
     This function add the operator for sequence pooling. 
-    This is applied on top of the input using pool_type mentioned
-    in the parameters.
+    It pools features of all time-steps of each instance, and is applied 
+    on top of the input using pool_type mentioned in the parameters. 
+
+    It supports four pool_type:
+
+    - average: :math:`Out[i] = \\frac{\sum_i X_i}{N}`
+    - sum:     :math:`Out[i] = \sum_jX_{ij}`
+    - sqrt:    :math:`Out[i] = \\frac{\sum_jX_{ij}}{\sqrt{len(X_i)}}`
+    - max:     :math:`Out[i] = max(X_i)`
+
+    .. code-block:: text
+
+       x is a 1-level LoDTensor:
+         x.lod = [[0, 2, 5, 7]]
+         x.data = [1, 3, 2, 4, 6, 5, 1]
+         x.dims = [7, 1]
+
+       then output is a Tensor:
+         out.dim = [3, 1]
+         with condition len(x.lod[-1]) - 1 == out.dims[0]
+
+       for different pool_type:
+         average: out.data = [2, 4, 3], where 2=(1+3)/2, 4=(2+4+6)/3, 3=(5+1)/2
+         sum    : out.data = [4, 12, 6], where 4=1+3, 12=2+4+6, 6=5+1
+         sqrt   : out.data = [2.82, 6.93, 4.24], where 2.82=(1+3)/sqrt(2),
+                    6.93=(2+4+6)/sqrt(3), 4.24=(5+1)/sqrt(2)
+         max    : out.data = [3, 6, 5], where 3=max(1,3), 6=max(2,4,6), 5=max(5,1)
+         
+    Args:
+        input(variable): The input variable which is a LoDTensor.
+        pool_type (string): The pooling type of sequence_pool. 
+            It supports average, sum, sqrt and max.
+
+    Returns:
+        The sequence pooling variable which is a Tensor.
+
+    Examples:
+
+        .. code-block:: python
+             
+             x = fluid.layers.data(name='x', shape=[7, 1], 
+                              dtype='float32', lod_level=1)
+             avg_x = fluid.layers.sequence_pool(input=x, pool_type='average')
+             sum_x = fluid.layers.sequence_pool(input=x, pool_type='sum')
+             sqrt_x = fluid.layers.sequence_pool(input=x, pool_type='sqrt')
+             max_x = fluid.layers.sequence_pool(input=x, pool_type='max')
     """
     helper = LayerHelper('sequence_pool', input=input, **kwargs)
     dtype = helper.input_dtype()
@@ -593,6 +639,72 @@ def sequence_pool(input, pool_type, **kwargs):
     return pool_out
 
 
+def sequence_first_step(input, **kwargs):
+    """
+    This funciton get the first step of sequence.
+
+    .. code-block:: text
+
+       x is a 1-level LoDTensor:
+         x.lod = [[0, 2, 5, 7]]
+         x.data = [1, 3, 2, 4, 6, 5, 1]
+         x.dims = [7, 1]
+
+       then output is a Tensor:
+         out.dim = [3, 1]
+         with condition len(x.lod[-1]) - 1 == out.dims[0]
+         out.data = [1, 2, 5], where 1=first(1,3), 2=first(2,4,6), 5=first(5,1)
+         
+    Args:
+        input(variable): The input variable which is a LoDTensor.
+
+    Returns:
+        The sequence's first step variable which is a Tensor.
+
+    Examples:
+
+        .. code-block:: python
+             
+             x = fluid.layers.data(name='x', shape=[7, 1], 
+                              dtype='float32', lod_level=1)
+             x_first_step = fluid.layers.sequence_first_step(input=x)
+    """
+    return sequence_pool(input=input, pool_type="first")
+
+
+def sequence_last_step(input, **kwargs):
+    """
+    This funciton get the last step of sequence.
+
+    .. code-block:: text
+
+       x is a 1-level LoDTensor:
+         x.lod = [[0, 2, 5, 7]]
+         x.data = [1, 3, 2, 4, 6, 5, 1]
+         x.dims = [7, 1]
+
+       then output is a Tensor:
+         out.dim = [3, 1]
+         with condition len(x.lod[-1]) - 1 == out.dims[0]
+         out.data = [3, 6, 1], where 3=last(1,3), 6=last(2,4,6), 1=last(5,1)
+         
+    Args:
+        input(variable): The input variable which is a LoDTensor.
+
+    Returns:
+        The sequence's last step variable which is a Tensor.
+
+    Examples:
+
+        .. code-block:: python
+             
+             x = fluid.layers.data(name='x', shape=[7, 1], 
+                              dtype='float32', lod_level=1)
+             x_last_step = fluid.layers.sequence_last_step(input=x)
+    """
+    return sequence_pool(input=input, pool_type="last")
+
+
 def pool2d(input,
            pool_size,
            pool_type,
@@ -1045,3 +1157,47 @@ def reduce_sum(input, dim=None, keep_dim=False):
             'reduce_all': True if dim == None else False
         })
     return out
+
+
+def reduce_mean(input, dim=None, keep_dim=False):
+    """
+    Computes the mean of tensor elements over the given dimension. 
+
+    Args:
+        input (Variable): The input variable which is a Tensor or LoDTensor.
+        dim (int|None): The dimension along which the mean is computed. If 
+            :attr:`None`, compute the mean over all elements of :attr:`input` 
+            and return a Tensor variable with a single element, otherwise 
+            must be in the range :math:`[-rank(input), rank(input))`. If 
+            :math:`dim < 0`, the dimension to reduce is :math:`rank + dim`.
+        keep_dim (bool): Whether to reserve the reduced dimension in the 
+            output Tensor. The result tensor will have one fewer dimension 
+            than the :attr:`input` unless :attr:`keep_dim` is true.
+
+    Returns:
+        Variable: The reduced Tensor variable.
+    
+    Examples:
+        .. code-block:: python
+
+            # x is a Tensor variable with following elements:
+            #    [[0.2, 0.3, 0.5, 0.9]
+            #     [0.1, 0.2, 0.6, 0.7]]
+            # Each example is followed by the correspending output tensor.
+            fluid.layers.reduce_mean(x)  # [0.4375]
+            fluid.layers.reduce_mean(x, dim=0)  # [0.15, 0.25, 0.55, 0.8]
+            fluid.layers.reduce_mean(x, dim=-1)  # [0.475, 0.4]
+            fluid.layers.reduce_mean(x, dim=1, keep_dim=True)  # [[0.475], [0.4]]
+    """
+    helper = LayerHelper('reduce_mean', **locals())
+    out = helper.create_tmp_variable(dtype=helper.input_dtype())
+    helper.append_op(
+        type='reduce_mean',
+        inputs={'X': input},
+        outputs={'Out': out},
+        attrs={
+            'dim': dim if dim != None else 0,
+            'keep_dim': keep_dim,
+            'reduce_all': True if dim == None else False
+        })
+    return out

python/paddle/v2/fluid/optimizer.py

@@ -207,7 +207,7 @@ class Optimizer(object):
 
         optimize_ops = self.create_optimization_pass(params_grads, loss,
                                                      startup_program)
-        return optimize_ops
+        return optimize_ops, params_grads
 
 
 class SGDOptimizer(Optimizer):

python/paddle/v2/fluid/param_attr.py

@@ -58,7 +58,9 @@ class ParamAttr(object):
     def to_kwargs(self, with_initializer=False):
         kwargs = {
             'name': self.name,
-            'learning_rate': self.learning_rate,
+            'optimize_attr': {
+                'learning_rate': self.learning_rate
+            },
             'regularizer': self.regularizer,
             'trainable': self.trainable,
             'clip_attr': self.clip

python/paddle/v2/fluid/profiler.py

 import paddle.v2.fluid.core as core
 from contextlib import contextmanager
+import os
 
 __all__ = ['CudaProfiler']
 
@@ -30,17 +31,21 @@ def cuda_profiler(output_file, output_mode=None, config=None):
             written into this file.
         output_mode (string) : The output mode has Key-Value pair format and
             Comma separated values format. It should be 'kvp' or 'csv'.
-        config (string) : The profiler options and counters can refer to
-            "Compute Command Line Profiler User Guide".
+        config (list of string) : The profiler options and counters can refer
+            to "Compute Command Line Profiler User Guide".
     """
     if output_mode is None:
         output_mode = 'csv'
     if output_mode not in ['kvp', 'csv']:
         raise ValueError("The output mode must be 'kvp' or 'csv'.")
     config = NVPROF_CONFIG if config is None else config
-    core.nvprof_init(output_file, output_mode, config)
+    config_file = 'nvprof_config_file'
+    with open(config_file, 'wb') as fp:
+        fp.writelines(["%s\n" % item for item in config])
+    core.nvprof_init(output_file, output_mode, config_file)
     # Enables profiler collection by the active CUDA profiling tool.
     core.nvprof_start()
     yield
     # Disables profiler collection.
     core.nvprof_stop()
+    os.remove(config_file)

python/paddle/v2/fluid/registry.py

@@ -8,7 +8,7 @@ import proto.framework_pb2 as framework_pb2
 from framework import OpProtoHolder, Variable, Program, Operator
 from paddle.v2.fluid.layer_helper import LayerHelper, unique_name
 
-__all__ = ['deprecated', 'register_layer']
+__all__ = ['deprecated', 'register_layer', 'autodoc']
 
 
 def _convert_(name):
@@ -175,12 +175,18 @@ def deprecated(func_or_class):
         """
         Wrap func with deprecated warning
         """
-        warnings.simplefilter('always', DeprecationWarning)  #turn off filter
+        warnings.simplefilter('always', DeprecationWarning)  # turn off filter
         warnings.warn(
             "Call to deprecated function {}.".format(func.__name__),
             category=DeprecationWarning,
             stacklevel=2)
-        warnings.simplefilter('default', DeprecationWarning)  #reset filter
+        warnings.simplefilter('default', DeprecationWarning)  # reset filter
         return func(*args, **kwargs)
 
     return func_wrapper
+
+
+def autodoc(func):
+    func.__doc__ = _generate_doc_string_(OpProtoHolder.instance().get_op_proto(
+        func.__name__))
+    return func

python/paddle/v2/fluid/tests/book/notest_recognize_digits_conv_dist.py

+from __future__ import print_function
+import numpy as np
+import paddle.v2 as paddle
+import paddle.v2.fluid as fluid
+import os
+
+images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype='float32')
+label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+conv_pool_1 = fluid.nets.simple_img_conv_pool(
+    input=images,
+    filter_size=5,
+    num_filters=20,
+    pool_size=2,
+    pool_stride=2,
+    act="relu")
+conv_pool_2 = fluid.nets.simple_img_conv_pool(
+    input=conv_pool_1,
+    filter_size=5,
+    num_filters=50,
+    pool_size=2,
+    pool_stride=2,
+    act="relu")
+
+predict = fluid.layers.fc(input=conv_pool_2, size=10, act="softmax")
+cost = fluid.layers.cross_entropy(input=predict, label=label)
+avg_cost = fluid.layers.mean(x=cost)
+optimizer = fluid.optimizer.Adam(learning_rate=0.01)
+optimize_ops, params_grads = optimizer.minimize(avg_cost)
+
+accuracy = fluid.evaluator.Accuracy(input=predict, label=label)
+
+BATCH_SIZE = 50
+PASS_NUM = 3
+train_reader = paddle.batch(
+    paddle.reader.shuffle(
+        paddle.dataset.mnist.train(), buf_size=500),
+    batch_size=BATCH_SIZE)
+
+place = fluid.CPUPlace()
+exe = fluid.Executor(place)
+t = fluid.DistributeTranspiler()
+pserver_endpoints = os.getenv("PSERVERS")
+training_role = os.getenv("TRAINING_ROLE",
+                          "TRAINER")  # get the training role: trainer/pserver
+t.transpile(optimize_ops, params_grads, pservers=pserver_endpoints, trainers=1)
+
+if training_role == "PSERVER":
+    pserver_prog = t.get_pserver_program(pserver_endpoints, optimize_ops)
+    exe.run(fluid.default_startup_program())
+    exe.run(pserver_prog)
+elif training_role == "TRAINER":
+    feeder = fluid.DataFeeder(feed_list=[images, label], place=place)
+    exe.run(fluid.default_startup_program())
+
+    for pass_id in range(PASS_NUM):
+        accuracy.reset(exe)
+        for data in train_reader():
+            loss, acc = exe.run(fluid.default_main_program(),
+                                feed=feeder.feed(data),
+                                fetch_list=[avg_cost] + accuracy.metrics)
+            pass_acc = accuracy.eval(exe)
+            # print loss, acc
+            if loss < 10.0 and pass_acc > 0.9:
+                # if avg cost less than 10.0 and accuracy is larger than 0.9, we think our code is good.
+                exit(0)
+
+        pass_acc = accuracy.eval(exe)
+        print("pass_id=" + str(pass_id) + " pass_acc=" + str(pass_acc))
+else:
+    print("environment var TRAINER_ROLE should be TRAINER os PSERVER")
+
+exit(1)

python/paddle/v2/fluid/tests/book/test_machine_translation.py

@@ -33,7 +33,7 @@ def encoder_decoder():
 
     fc1 = fluid.layers.fc(input=src_embedding, size=hidden_dim * 4, act='tanh')
     lstm_hidden0, lstm_0 = layers.dynamic_lstm(input=fc1, size=hidden_dim * 4)
-    encoder_out = layers.sequence_pool(input=lstm_hidden0, pool_type="last")
+    encoder_out = layers.sequence_last_step(input=lstm_hidden0)
 
     # decoder
     trg_language_word = layers.data(

python/paddle/v2/fluid/tests/book/test_recommender_system.py

@@ -125,10 +125,11 @@ def model():
 
     # need cos sim
     inference = layers.cos_sim(X=usr_combined_features, Y=mov_combined_features)
+    scale_infer = layers.scale(x=inference, scale=5.0)
 
     label = layers.data(name='score', shape=[1], dtype='float32')
 
-    square_cost = layers.square_error_cost(input=inference, label=label)
+    square_cost = layers.square_error_cost(input=scale_infer, label=label)
 
     avg_cost = layers.mean(x=square_cost)
 
@@ -141,7 +142,7 @@ def main():
     opts = sgd_optimizer.minimize(cost)
 
     if USE_GPU:
-        place = core.GPUPlace(0)
+        place = core.CUDAPlace(0)
     else:
         place = core.CPUPlace()
 

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

@@ -90,12 +90,10 @@ def get_numeric_gradient(scope,
     def product(dim):
         return reduce(lambda a, b: a * b, dim, 1)
 
-    ctx = core.DeviceContext.create(core.CPUPlace())
-
     def get_output():
         sum = []
         for output_name in output_names:
-            op.run(scope, ctx)
+            op.run(scope, core.CPUPlace())
             sum.append(
                 np.array(scope.find_var(output_name).get_tensor()).mean())
         return np.array(sum).mean()
@@ -318,7 +316,7 @@ class OpTest(unittest.TestCase):
     def check_output(self, atol=1e-5):
         places = [core.CPUPlace()]
         if core.is_compile_gpu() and core.op_support_gpu(self.op_type):
-            places.append(core.GPUPlace(0))
+            places.append(core.CUDAPlace(0))
         for place in places:
             self.check_output_with_place(place, atol)
 
@@ -381,7 +379,7 @@ class OpTest(unittest.TestCase):
                                "Gradient Check On %s" % str(cpu_place))
 
         if core.is_compile_gpu() and self.op.support_gpu():
-            gpu_place = core.GPUPlace(0)
+            gpu_place = core.CUDAPlace(0)
             gpu_analytic_grads = self._get_gradient(inputs_to_check, gpu_place,
                                                     output_names, no_grad_set)
 

python/paddle/v2/fluid/tests/test_adagrad_op.py

@@ -113,8 +113,7 @@ class TestSparseAdagradOp(unittest.TestCase):
             LearningRate='LearningRate',
             epsilon=2.0)
 
-        ctx = core.DeviceContext.create(place)
-        adagrad_op.run(scope, ctx)
+        adagrad_op.run(scope, place)
 
         # get and compare moment result
         moment_result_array = np.array(moment)
@@ -168,7 +167,7 @@ class TestSparseAdagradOp(unittest.TestCase):
     def test_sparse_adagrad(self):
         places = [core.CPUPlace()]
         if core.is_compile_gpu():
-            places.append(core.GPUPlace(0))
+            places.append(core.CUDAPlace(0))
         for place in places:
             self.check_with_place(place)
 

python/paddle/v2/fluid/tests/test_batch_norm_op.py

@@ -208,7 +208,7 @@ class TestBatchNormOp(OpTest):
         print 'python: NHWC, NCHW, backward checking passed'
 
     def test_forward_backward(self):
-        def test_with_place(place, tensor_format, shape):
+        def test_with_place(place, data_layout, shape):
             # attr
             epsilon = 0.00001
             momentum = 0.9
@@ -292,12 +292,11 @@ class TestBatchNormOp(OpTest):
                 SavedVariance="saved_variance",
                 # attrs
                 is_test=False,
-                tensor_format=tensor_format,
+                data_layout=data_layout,
                 momentum=momentum,
                 epsilon=epsilon)
 
-            ctx = core.DeviceContext.create(place)
-            batch_norm_op.run(scope, ctx)
+            batch_norm_op.run(scope, place)
 
             # check forward result
             self.__assert_close(y_tensor, y_out, "y_out")
@@ -305,13 +304,13 @@ class TestBatchNormOp(OpTest):
             self.__assert_close(saved_variance_tensor, saved_variance,
                                 "saved_variance")
             self.__assert_close(mean_out_tensor, mean_out, "mean_out")
-            if isinstance(place, core.GPUPlace):
+            if isinstance(place, core.CUDAPlace):
                 atol = 5e-2
             else:
                 atol = 1e-4
             self.__assert_close(variance_out_tensor, variance_out,
                                 "variance_out", atol)
-            print "op test forward passed: ", str(place), tensor_format
+            print "op test forward passed: ", str(place), data_layout
 
             # run backward
             batch_norm_op_grad = get_backward_op(scope, batch_norm_op, set())
@@ -320,7 +319,7 @@ class TestBatchNormOp(OpTest):
                 ["y_out", "mean", "variance", "saved_mean", "saved_variance"],
                 place,
                 feed_dict={"y_out": y_grad})
-            batch_norm_op_grad.run(scope, ctx)
+            batch_norm_op_grad.run(scope, place)
 
             x_grad_tensor = create_or_get_tensor(scope,
                                                  grad_var_name("x_val"), None,
@@ -336,11 +335,15 @@ class TestBatchNormOp(OpTest):
             self.__assert_close(x_grad_tensor, x_grad_ref, "x_grad")
             self.__assert_close(scale_grad_tensor, scale_grad_ref, "scale_grad")
             self.__assert_close(bias_grad_tensor, bias_grad_ref, "bias_grad")
-            print "op test backward passed: ", str(place), tensor_format
+            print "op test backward passed: ", str(place), data_layout
 
         places = [core.CPUPlace()]
         if core.is_compile_gpu() and core.op_support_gpu("batch_norm"):
-            places.append(core.GPUPlace(0))
+            places.append(core.CUDAPlace(0))
+
+            core.init_devices(["CPU", "GPU:0"])
+        else:
+            core.init_devices(["CPU"])
         for place in places:
             for data_format in ["NCHW", "NHWC"]:
                 test_with_place(place, data_format, [2, 3, 4, 5])

python/paddle/v2/fluid/tests/test_beam_search_decode_op.py

@@ -57,8 +57,7 @@ class TestBeamSearchDecodeOp(unittest.TestCase):
             SentenceIds="sentence_ids",
             SentenceScores="sentence_scores")
 
-        ctx = core.DeviceContext.create(self.cpu_place)
-        beam_search_decode_op.run(self.scope, ctx)
+        beam_search_decode_op.run(self.scope, self.cpu_place)
 
         expected_lod = [[0, 4, 8], [0, 1, 3, 6, 9, 10, 13, 16, 19]]
         self.assertEqual(sentence_ids.lod(), expected_lod)

python/paddle/v2/fluid/tests/test_beam_search_op.py

@@ -14,7 +14,6 @@ def create_tensor(scope, name, np_data):
 class BeamSearchOpTester(unittest.TestCase):
     def setUp(self):
         self.scope = core.Scope()
-        self.ctx = core.DeviceContext.create(core.CPUPlace())
         self._create_ids()
         self._create_scores()
         self._create_pre_ids()
@@ -32,7 +31,7 @@ class BeamSearchOpTester(unittest.TestCase):
             level=0,
             beam_size=2,
             end_id=0, )
-        op.run(self.scope, self.ctx)
+        op.run(self.scope, core.CPUPlace())
         selected_ids = self.scope.find_var("selected_ids").get_tensor()
         print 'selected_ids', np.array(selected_ids)
         print 'lod', selected_ids.lod()

python/paddle/v2/fluid/tests/test_cond_op.py

@@ -65,8 +65,7 @@ class TestCondOp(unittest.TestCase):
         self.create_global_variables()
         self.create_cond_op()
         self.create_sub_net()
-        ctx = core.DeviceContext.create(core.CPUPlace())
-        self.condop.run(self.scope, ctx)
+        self.condop.run(self.scope, core.CPUPlace())
         return np.array(self.scope.find_var("Out").get_tensor())
 
     def create_global_variables(self):

python/paddle/v2/fluid/tests/test_dyn_rnn.py

@@ -63,8 +63,7 @@ class TestDynRNN(unittest.TestCase):
 
             all_timesteps = fluid.layers.array_to_lod_tensor(
                 x=out, table=rank_table)
-            last = fluid.layers.sequence_pool(
-                input=all_timesteps, pool_type='last')
+            last = fluid.layers.sequence_last_step(input=all_timesteps)
             logits = fluid.layers.fc(input=last, size=1, act=None)
             loss = fluid.layers.sigmoid_cross_entropy_with_logits(
                 x=logits, label=label)
@@ -101,7 +100,7 @@ class TestDynRNN(unittest.TestCase):
                 rnn.update_memory(mem, out_)
                 rnn.output(out_)
 
-            last = fluid.layers.sequence_pool(input=rnn(), pool_type='last')
+            last = fluid.layers.sequence_last_step(input=rnn())
             logits = fluid.layers.fc(input=last, size=1, act=None)
             label = fluid.layers.data(name='label', shape=[1], dtype='float32')
             loss = fluid.layers.sigmoid_cross_entropy_with_logits(

python/paddle/v2/fluid/tests/test_gaussian_random_op.py

 import unittest
+import numpy
+
+import paddle.v2.fluid as fluid
 import paddle.v2.fluid.core as core
 from paddle.v2.fluid.op import Operator
-import numpy
+from paddle.v2.fluid.executor import Executor
 
 
 class TestGaussianRandomOp(unittest.TestCase):
+    def setUp(self):
+        self.op_type = "gaussian_random"
+        self.inputs = {}
+        self.attrs = {"shape": [1000, 784], "mean": .0, "std": 1., "seed": 10}
+
+        self.outputs = ["Out"]
+
     def test_cpu(self):
-        self.gaussian_random_test(place=core.CPUPlace())
+        self.gaussian_random_test(place=fluid.CPUPlace())
 
     def test_gpu(self):
         if core.is_compile_gpu():
-            self.gaussian_random_test(place=core.GPUPlace(0))
+            self.gaussian_random_test(place=fluid.CUDAPlace(0))
 
     def gaussian_random_test(self, place):
-        scope = core.Scope()
-        scope.var('Out').get_tensor()
-
-        op = Operator(
-            "gaussian_random",
-            Out='Out',
-            shape=[1000, 784],
-            mean=.0,
-            std=1.,
-            seed=10)
-
-        context = core.DeviceContext.create(place)
-        op.run(scope, context)
-        tensor = numpy.array(scope.find_var('Out').get_tensor())
+
+        program = fluid.Program()
+        block = program.global_block()
+        vout = block.create_var(name="Out")
+        op = block.append_op(
+            type=self.op_type, outputs={"Out": vout}, attrs=self.attrs)
+
+        op.desc.infer_var_type(block.desc)
+        op.desc.infer_shape(block.desc)
+
+        fetch_list = []
+        for var_name in self.outputs:
+            fetch_list.append(block.var(var_name))
+
+        exe = Executor(place)
+        outs = exe.run(program, fetch_list=fetch_list)
+        tensor = outs[0]
+
         self.assertAlmostEqual(numpy.mean(tensor), .0, delta=0.1)
         self.assertAlmostEqual(numpy.std(tensor), 1., delta=0.1)
 

python/paddle/v2/fluid/tests/test_is_empty_op.py

@@ -33,8 +33,7 @@ class TestIsEmptyOp(unittest.TestCase):
 
     def one_case(self, input, target):
         op = Operator(type="is_empty", X=input, Out="out")
-        ctx = core.DeviceContext.create(core.CPUPlace())
-        op.run(self.scope, ctx)
+        op.run(self.scope, core.CPUPlace())
         out = self.scope.var("out").get_tensor()
         self.assertEqual(np.array(out)[0], target)
 

python/paddle/v2/fluid/tests/test_optimizer.py

@@ -27,7 +27,7 @@ class TestOptimizer(unittest.TestCase):
         block.append_op(
             type="mean", inputs={"X": mul_out}, outputs={"Out": mean_out})
         sgd_optimizer = optimizer.SGDOptimizer(learning_rate=0.01)
-        opts = sgd_optimizer.minimize(mean_out, init_program)
+        opts, _ = sgd_optimizer.minimize(mean_out, init_program)
         self.assertEqual(len(opts), 1)
         sgd_op = opts[0]
         self.assertEqual(sgd_op.type, "sgd")
@@ -57,7 +57,7 @@ class TestOptimizer(unittest.TestCase):
         learning_rate = 0.01
         sgd_optimizer = optimizer.SGDOptimizer(
             learning_rate=learning_rate, global_step=global_step)
-        opts = sgd_optimizer.minimize(mean_out, init_program)
+        opts, _ = sgd_optimizer.minimize(mean_out, init_program)
         self.assertEqual(len(opts), 2)
         sgd_op = opts[0]
         self.assertEqual(sgd_op.type, "sgd")

python/paddle/v2/fluid/tests/test_profiler.py

@@ -3,6 +3,7 @@ import numpy as np
 import paddle.v2.fluid as fluid
 import paddle.v2.fluid.profiler as profiler
 import paddle.v2.fluid.layers as layers
+import os
 
 
 class TestProfiler(unittest.TestCase):
@@ -14,14 +15,16 @@ class TestProfiler(unittest.TestCase):
         data = layers.data(name='data', shape=[3, 28, 28], dtype='float32')
         conv = layers.conv2d(data, 20, 3, stride=[1, 1], padding=[1, 1])
 
-        place = fluid.GPUPlace(0)
+        place = fluid.CUDAPlace(0)
         exe = fluid.Executor(place)
         exe.run(fluid.default_startup_program())
 
-        with profiler.cuda_profiler("cuda_profiler.txt", 'csv') as nvprof:
+        output_file = 'cuda_profiler.txt'
+        with profiler.cuda_profiler(output_file, 'csv') as nvprof:
             for i in range(epoc):
                 input = np.random.random(dshape).astype('float32')
                 exe.run(fluid.default_main_program(), feed={'data': input})
+        os.remove(output_file)
 
 
 if __name__ == '__main__':

python/paddle/v2/fluid/tests/test_reorder_lod_tensor.py

+import unittest
+import paddle.v2.fluid as fluid
+import numpy
+
+
+class TestReorderLoDTensor(unittest.TestCase):
+    def test_reorder(self):
+        dat = fluid.layers.data(name='input', shape=[1], lod_level=2)
+        dat.stop_gradient = False
+        rank_dat = fluid.layers.data(name='ref', shape=[1], lod_level=1)
+        table = fluid.layers.lod_rank_table(rank_dat)
+        new_dat = fluid.layers.reorder_lod_tensor_by_rank(
+            x=dat, rank_table=table)
+        loss = fluid.layers.mean(x=new_dat)
+        fluid.backward.append_backward_ops(loss=loss)
+
+        cpu = fluid.CPUPlace()
+        exe = fluid.Executor(cpu)
+        exe.run(fluid.default_startup_program())
+
+        ref = fluid.Tensor()
+        ref_lod = [0, 3, 4, 7, 8, 14]
+        ref.set_lod([ref_lod])
+
+        ref.set(numpy.random.random(size=[14, 1]).astype('float32'), cpu)
+        input = fluid.Tensor()
+        lod_level_0 = numpy.random.randint(low=1, high=5, size=5)
+        lod_level_0 = [0] + numpy.cumsum(lod_level_0).tolist()
+        lod_level_1 = numpy.random.randint(low=1, high=5, size=lod_level_0[-1])
+        lod_level_1 = [0] + numpy.cumsum(lod_level_1).tolist()
+
+        input.set_lod([lod_level_0, lod_level_1])
+        input.set(
+            numpy.random.random(size=[lod_level_1[-1], 1]).astype('float32'),
+            cpu)
+
+        ig = exe.run(fluid.default_main_program(),
+                     feed={'input': input,
+                           'ref': ref},
+                     fetch_list=['input@GRAD'],
+                     return_numpy=False)[0]
+        self.assertAlmostEqual(numpy.array(ig).sum(), 1.0, delta=0.001)
+        self.assertEqual(input.lod(), ig.lod())
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/v2/fluid/tests/test_sgd_op.py

@@ -55,8 +55,7 @@ class TestSparseSGDOp(unittest.TestCase):
             Grad='Grad',
             ParamOut='Param',
             LearningRate='LearningRate')
-        ctx = core.DeviceContext.create(place)
-        sgd_op.run(scope, ctx)
+        sgd_op.run(scope, place)
 
         # get and compare result
         result_array = np.array(param)
@@ -79,7 +78,7 @@ class TestSparseSGDOp(unittest.TestCase):
     def test_sparse_sgd(self):
         places = [core.CPUPlace()]
         if core.is_compile_gpu():
-            places.append(core.GPUPlace(0))
+            places.append(core.CUDAPlace(0))
         for place in places:
             self.check_with_place(place)
 

python/paddle/v2/fluid/tests/test_uniform_random_op.py

 import unittest
+import numpy
+
 from paddle.v2.fluid.op import Operator
 import paddle.v2.fluid.core as core
-import numpy
+import paddle.v2.fluid as fluid
 
 
 class TestUniformRandomOp(unittest.TestCase):
-    def test_uniform_random_cpu(self):
+    def setUp(self):
+        self.op_type = "uniform_random"
+        self.inputs = {}
+        self.attrs = {
+            "shape": [1000, 784],
+            "min": -5.0,
+            "max": 10.0,
+            "seed": 10
+        }
+        self.outputs = ["Out"]
+
+    def test_cpu(self):
         self.uniform_random_test(place=core.CPUPlace())
 
-    def test_uniform_random_gpu(self):
+    def test_gpu(self):
         if core.is_compile_gpu():
-            self.uniform_random_test(place=core.GPUPlace(0))
+            self.uniform_random_test(place=core.CUDAPlace(0))
 
     def uniform_random_test(self, place):
-        scope = core.Scope()
-        scope.var('X').get_tensor()
-
-        op = Operator(
-            "uniform_random",
-            Out='X',
-            shape=[1000, 784],
-            min=-5.0,
-            max=10.0,
-            seed=10)
-
-        ctx = core.DeviceContext.create(place)
-        op.run(scope, ctx)
-        tensor = numpy.array(scope.find_var('X').get_tensor())
+        program = fluid.Program()
+        block = program.global_block()
+        vout = block.create_var(name="Out")
+        op = block.append_op(
+            type=self.op_type, outputs={"Out": vout}, attrs=self.attrs)
+
+        op.desc.infer_var_type(block.desc)
+        op.desc.infer_shape(block.desc)
+
+        fetch_list = []
+        for var_name in self.outputs:
+            fetch_list.append(block.var(var_name))
+
+        exe = fluid.Executor(place)
+        outs = exe.run(program, fetch_list=fetch_list)
+
+        tensor = outs[0]
+
         self.assertAlmostEqual(tensor.mean(), 2.5, delta=0.1)
 
 

python/setup.py.in

@@ -79,8 +79,7 @@ if '${CMAKE_SYSTEM_PROCESSOR}' not in ['arm', 'armv7-a', 'aarch64']:
 
 # the prefix is sys.prefix which should always be usr
 paddle_bin_dir = 'opt/paddle/bin'
-paddle_bins = ['${PADDLE_BINARY_DIR}/paddle/scripts/paddle_usage',
-               '${PADDLE_BINARY_DIR}/paddle/trainer/paddle_trainer',
+paddle_bins = ['${PADDLE_BINARY_DIR}/paddle/trainer/paddle_trainer',
                '${PADDLE_BINARY_DIR}/paddle/trainer/paddle_merge_model',
                '${PADDLE_BINARY_DIR}/paddle/pserver/paddle_pserver_main',
                '${PADDLE_BINARY_DIR}/paddle/scripts/paddle']