Merge remote-tracking branch 'ups/develop' into refine/im2col

AUTHORS.md

@@ -46,6 +46,7 @@
 | tianbingsz | Tian-Bing Xu |
 | tpatejko | Tomasz Patejko |
 | typhoonzero | Yi Wu |
+| velconia | Qi-Yang Min |
 | wanghaoshuang | Hao-Shuang Wang |
 | wangyang59 | Yang Wang |
 | wangzhen-nlp | Zhen Wang |

benchmark/fluid/fluid_benchmark.py

@@ -85,8 +85,7 @@ def dist_transpile(trainer_id, args):
         trainer_id,
         pservers=pserver_endpoints,
         trainers=trainers,
-        sync_mode=not args.async_mode,
-        slice_var_up=not args.no_split_var)
+        sync_mode=not args.async_mode)
     if training_role == "PSERVER":
         pserver_program = t.get_pserver_program(current_endpoint)
         pserver_startup_program = t.get_startup_program(current_endpoint,

cmake/external/grpc.cmake

@@ -50,7 +50,7 @@ ExternalProject_Add(
     UPDATE_COMMAND  ""
     CONFIGURE_COMMAND ""
     BUILD_IN_SOURCE 1
-    PATCH_COMMAND git apply ${PADDLE_SOURCE_DIR}/patches/grpc/fix_too_early_destory.patch
+    PATCH_COMMAND cp ${PADDLE_SOURCE_DIR}/patches/grpc/grpc_library.h ${GRPC_SOURCES_DIR}/src/extern_grpc/include/grpcpp/impl/codegen/grpc_library.h && cp ${PADDLE_SOURCE_DIR}/patches/grpc/completion_queue.h ${GRPC_SOURCES_DIR}/src/extern_grpc/include/grpcpp/impl/codegen/completion_queue.h
     # NOTE(yuyang18):
     # Disable -Werror, otherwise the compile will fail in MacOS.
     # It seems that we cannot configure that by make command.

cmake/inference_lib.cmake

@@ -148,18 +148,11 @@ if (WITH_ANAKIN AND WITH_GPU)
      list(APPEND inference_deps anakin_inference_lib)
 endif()
 
-copy(inference_api_lib DEPS paddle_inference_api paddle_inference_api_shared
-  SRCS ${src_dir}/${module}/paddle_inference_api.h 
-       ${src_dir}/${module}/demo_ci
-       ${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/libpaddle_inference_api*
-  DSTS ${dst_dir}/inference ${dst_dir}/inference ${dst_dir}/inference
-)
-list(APPEND inference_deps inference_api_lib)
-
 set(module "inference")
 copy(inference_lib DEPS ${inference_deps}
   SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.*
-  DSTS ${dst_dir}/${module} ${dst_dir}/${module}
+       ${src_dir}/${module}/api/paddle_inference_api.h ${src_dir}/${module}/api/demo_ci
+  DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module}
 )
 
 set(module "platform")

paddle/fluid/framework/CMakeLists.txt

@@ -8,9 +8,9 @@ cc_test(ddim_test SRCS ddim_test.cc DEPS ddim)
 nv_test(dim_test SRCS dim_test.cu DEPS ddim)
 cc_library(data_type SRCS data_type.cc DEPS framework_proto ddim device_context)
 if(WITH_GPU)
-  nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS place memory data_type)
+  nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS place memory data_type device_context)
 else()
-  cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS place memory data_type)
+  cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS place memory data_type device_context)
 endif()
 
 cc_test(tensor_test SRCS tensor_test.cc DEPS tensor)
@@ -110,7 +110,7 @@ cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows)
 
 cc_test(op_kernel_type_test SRCS op_kernel_type_test.cc DEPS place device_context framework_proto)
 cc_test(cow_ptr_tests SRCS details/cow_ptr_test.cc)
-      
+
 # cc_test(channel_test SRCS channel_test.cc)
 cc_test(tuple_test SRCS tuple_test.cc )
 

paddle/fluid/framework/details/multi_devices_graph_builder.cc

@@ -715,6 +715,7 @@ void MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result,
       result->CreateOpNode(node->Op()), *node->Op(), local_scopes_[op_dev_id],
       node->Op()->Type(), places_[op_dev_id]));
 
+  // TODO(panyx0718): This might not be needed anymore.
   if (node->Op()->Type() == "send_barrier") {
     ConnectOp(result, result->Get<GraphOps>("ops").back().get(), "send");
   } else if (node->Op()->Type() == "recv") {

paddle/fluid/framework/ir/graph.cc

@@ -24,6 +24,68 @@ namespace paddle {
 namespace framework {
 namespace ir {
 
+std::vector<std::string> FindDistTrainSendVars(
+    const std::vector<ir::Node *> &nodes) {
+  std::vector<std::string> send_vars;
+  // since parameters are all in block 0,
+  // it's enough to only scan send ops in block 0
+  for (auto &node : nodes) {
+    auto op_vars = node->Op()->InputArgumentNames();
+    send_vars.reserve(send_vars.size() +
+                      std::distance(op_vars.begin(), op_vars.end()));
+    send_vars.insert(send_vars.end(), op_vars.begin(), op_vars.end());
+  }
+  return send_vars;
+}
+
+std::vector<std::string> FindDistTrainRecvVars(
+    const std::vector<ir::Node *> &nodes) {
+  std::vector<std::string> recv_vars;
+  for (auto &node : nodes) {
+    auto op_vars = node->Op()->OutputArgumentNames();
+    recv_vars.reserve(recv_vars.size() +
+                      std::distance(op_vars.begin(), op_vars.end()));
+    recv_vars.insert(recv_vars.end(), op_vars.begin(), op_vars.end());
+  }
+  return recv_vars;
+}
+
+bool IsDistTrainOp(ir::Node *node, const std::vector<std::string> &send_vars,
+                   const std::vector<std::string> &recv_vars) {
+  if (send_vars.size() == 0 || recv_vars.size() == 0) {
+    return false;
+  }
+
+  /**
+   * Check any of opvars contains `.block` and in sendvars
+   */
+  auto checker = [](const std::vector<std::string> &opvars,
+                    const std::vector<std::string> &rpc_vars) -> bool {
+    for (auto &var : opvars) {
+      // a variable name with the suffix `.block` means it's a splited
+      // variable by (DistributeTranspiler)
+      // [python/paddle/fluid/transpiler/distribute_transpiler.py]
+      if (var.find(".block") != std::string::npos &&
+          std::find(rpc_vars.begin(), rpc_vars.end(), var) != rpc_vars.end()) {
+        return true;
+      }
+    }
+    return false;
+  };
+
+  std::vector<std::string> input_var_names;
+  std::vector<std::string> output_var_names;
+  for (ir::Node *input : node->inputs) {
+    input_var_names.push_back(input->Name());
+  }
+  for (ir::Node *output : node->outputs) {
+    output_var_names.push_back(output->Name());
+  }
+
+  return checker(output_var_names, send_vars) ||
+         checker(input_var_names, recv_vars);
+}
+
 Graph::Graph(const ProgramDesc &program) : program_(program) {
   VLOG(3) << "block in program:" << program_.Size();
   std::unordered_map<std::string, VarDesc *> all_vars;
@@ -61,6 +123,64 @@ Graph::Graph(const ProgramDesc &program) : program_(program) {
       var->inputs.push_back(node);
     }
   }
+
+  std::vector<ir::Node *> send_ops;
+  ir::Node *send_bar = nullptr;
+  std::vector<ir::Node *> recv_ops;
+  ir::Node *fetch_bar = nullptr;
+  for (ir::Node *node : Nodes()) {
+    if (node->Name() == "send") {
+      send_ops.push_back(node);
+    } else if (node->Name() == "send_barrier") {
+      PADDLE_ENFORCE(!send_bar, "only has one send barrier");
+      send_bar = node;
+    } else if (node->Name() == "recv") {
+      recv_ops.push_back(node);
+    } else if (node->Name() == "fetch_barrier") {
+      PADDLE_ENFORCE(!fetch_bar, "only has one fetch barrier");
+      fetch_bar = node;
+    }
+  }
+  if (send_bar) {
+    for (ir::Node *send : send_ops) {
+      ir::Node *dep_var = CreateControlDepVar();
+      send->outputs.push_back(dep_var);
+      dep_var->inputs.push_back(send);
+      send_bar->inputs.push_back(dep_var);
+      dep_var->outputs.push_back(send_bar);
+    }
+    for (ir::Node *recv : recv_ops) {
+      ir::Node *dep_var = CreateControlDepVar();
+      recv->inputs.push_back(dep_var);
+      dep_var->outputs.push_back(recv);
+      send_bar->outputs.push_back(dep_var);
+      dep_var->inputs.push_back(send_bar);
+    }
+  }
+  if (fetch_bar) {
+    for (ir::Node *recv : recv_ops) {
+      ir::Node *dep_var = CreateControlDepVar();
+      recv->outputs.push_back(dep_var);
+      dep_var->inputs.push_back(recv);
+      fetch_bar->inputs.push_back(dep_var);
+      dep_var->outputs.push_back(fetch_bar);
+    }
+  }
+
+  std::vector<std::string> send_vars = FindDistTrainSendVars(send_ops);
+  std::vector<std::string> recv_vars = FindDistTrainRecvVars(recv_ops);
+  for (ir::Node *node : Nodes()) {
+    if (IsDistTrainOp(node, send_vars, recv_vars)) {
+      if (fetch_bar && node->Name() == "concat") {
+        ir::Node *dep_var = CreateControlDepVar();
+        fetch_bar->outputs.push_back(dep_var);
+        dep_var->inputs.push_back(fetch_bar);
+        node->inputs.push_back(dep_var);
+        dep_var->outputs.push_back(node);
+      }
+    }
+  }
+
   /**
    * We only handle write after read(WAR), since it should not have a write
    * after write in program. If there are write after write operators, we need

paddle/fluid/inference/CMakeLists.txt

@@ -14,8 +14,15 @@ cc_library(paddle_fluid_api
 
 get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES)
 
+# paddle_fluid_origin exclude inference api interface
+cc_library(paddle_fluid_origin DEPS ${fluid_modules} paddle_fluid_api)
+
+if(NOT APPLE)
+  add_subdirectory(api)
+endif()
+
 # Create static library
-cc_library(paddle_fluid DEPS ${fluid_modules} paddle_fluid_api)
+cc_library(paddle_fluid DEPS ${fluid_modules} paddle_fluid_api paddle_inference_api)
 if(NOT APPLE)
   # TODO(liuyiqu: Temporarily disable the link flag because it is not support on Mac.
   set(LINK_FLAGS "-Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/paddle_fluid.sym")
@@ -24,7 +31,7 @@ endif()
 
 # Create shared library
 cc_library(paddle_fluid_shared SHARED
-    SRCS io.cc
+    SRCS io.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api_impl.cc
     DEPS ${fluid_modules} paddle_fluid_api)
 
 set_target_properties(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid)
@@ -32,12 +39,21 @@ if(NOT APPLE)
   # TODO(liuyiqun): Temporarily disable the link flag because it is not support on Mac.
   set(LINK_FLAGS "-Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/paddle_fluid.map")
   set_target_properties(paddle_fluid_shared PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
+  # check symbol hidden
+  FILE(WRITE ${CMAKE_CURRENT_BINARY_DIR}/check_symbol.cmake
+    "execute_process(COMMAND bash -c \"${CMAKE_CURRENT_SOURCE_DIR}/check_symbol.sh"
+    " ${CMAKE_CURRENT_BINARY_DIR}/libpaddle_fluid.so\" RESULT_VARIABLE symbol_res)\n"
+    "if(NOT \"\${symbol_res}\" STREQUAL \"0\")\n"
+    "  message(FATAL_ERROR \"Check symbol failed.\")\n"
+    "endif()\n")
+  add_custom_command(
+    OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/.check_symbol"
+    COMMAND ${CMAKE_COMMAND} -P "${CMAKE_CURRENT_BINARY_DIR}/check_symbol.cmake"
+    DEPENDS paddle_fluid_shared)
+  add_custom_target(check_symbol ALL DEPENDS "${CMAKE_CURRENT_BINARY_DIR}/.check_symbol")
 endif()
 
 if(WITH_TESTING)
   # both tests/book and analysis depends the models that generated by python/paddle/fluid/tests/book
   add_subdirectory(tests/book)
 endif()
-if(NOT APPLE)
-  add_subdirectory(api)
-endif()

paddle/fluid/inference/api/CMakeLists.txt

@@ -42,35 +42,8 @@ function(inference_api_test TARGET_NAME)
     endif(WITH_TESTING)
 endfunction(inference_api_test)
 
-cc_library(paddle_inference_api
-    SRCS api.cc api_impl.cc
-    DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
-if(NOT APPLE)
-  set(LINK_FLAGS "-Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/api.sym")
-  set_target_properties(paddle_inference_api PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
-endif()
-
-# Here the shared library doesn't depend on other fluid libraries, or double free will occur.
-cc_library(paddle_inference_api_shared SHARED
-    SRCS api.cc api_impl.cc)
-add_dependencies(paddle_inference_api_shared ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
-set_target_properties(paddle_inference_api_shared PROPERTIES OUTPUT_NAME paddle_inference_api)
+cc_library(paddle_inference_api SRCS api.cc api_impl.cc DEPS lod_tensor)
 
-if(NOT APPLE)
-  set(LINK_FLAGS "-Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/api.map")
-  set_target_properties(paddle_inference_api_shared PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
-  FILE(WRITE ${CMAKE_CURRENT_BINARY_DIR}/check_symbol.cmake
-    "execute_process(COMMAND bash -c \"${CMAKE_CURRENT_SOURCE_DIR}/check_symbol.sh"
-    " ${CMAKE_CURRENT_BINARY_DIR}/libpaddle_inference_api.so\" RESULT_VARIABLE symbol_res)\n"
-    "if(NOT \"\${symbol_res}\" STREQUAL \"0\")\n"
-    "  message(FATAL_ERROR \"Check symbol failed.\")\n"
-    "endif()\n")
-  add_custom_command(
-    OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/.check_symbol"
-    COMMAND ${CMAKE_COMMAND} -P "${CMAKE_CURRENT_BINARY_DIR}/check_symbol.cmake"
-    DEPENDS paddle_inference_api_shared)
-  add_custom_target(check_symbol ALL DEPENDS "${CMAKE_CURRENT_BINARY_DIR}/.check_symbol")
-endif()
 
 cc_test(test_paddle_inference_api
         SRCS api_tester.cc

paddle/fluid/inference/api/api.map

-{
-	global:
-		*paddle*;
-	local:
-		*;
-};

paddle/fluid/inference/api/api.sym

-*paddle*

paddle/fluid/inference/api/demo_ci/CMakeLists.txt

@@ -55,11 +55,9 @@ endif()
 # Note: libpaddle_inference_api.so/a must put before libpaddle_fluid.so/a
 if(WITH_STATIC_LIB)
   set(DEPS
-      ${PADDLE_LIB}/paddle/fluid/inference/libpaddle_inference_api.a
       ${PADDLE_LIB}/paddle/fluid/inference/libpaddle_fluid.a)
 else()
   set(DEPS
-      ${PADDLE_LIB}/paddle/fluid/inference/libpaddle_inference_api.so
       ${PADDLE_LIB}/paddle/fluid/inference/libpaddle_fluid.so)
 endif()
 set(EXTERNAL_LIB "-lrt -ldl -lpthread")

paddle/fluid/inference/api/check_symbol.sh → paddle/fluid/inference/check_symbol.sh

@@ -3,8 +3,8 @@
 lib=$1
 if [ $# -ne 1 ]; then echo "No input library"; exit -1 ; fi
 
-num_paddle_syms=$(nm -D --defined-only ${lib} | grep paddle | wc -l)
-num_google_syms=$(nm -D --defined-only ${lib} | grep google | wc -l)
+num_paddle_syms=$(nm -D ${lib} | grep paddle | wc -l)
+num_google_syms=$(nm -D ${lib} | grep google | grep -v paddle | grep T | wc -l)
 
 if [ $num_paddle_syms -le 0 ]; then echo "Have no paddle symbols"; exit -1 ; fi
 if [ $num_google_syms -ge 1 ]; then echo "Have some google symbols"; exit -1 ; fi

paddle/fluid/inference/tensorrt/convert/CMakeLists.txt

 # Add TRT tests
 nv_library(tensorrt_converter
-  SRCS mul_op.cc conv2d_op.cc fc_op.cc
-  DEPS tensorrt_engine mul_op)
+  SRCS mul_op.cc conv2d_op.cc fc_op.cc pool2d_op.cc
+  DEPS tensorrt_engine operator scope framework_proto op_registry)
 
 nv_test(test_op_converter SRCS test_op_converter.cc DEPS
   ${FLUID_CORE_MODULES} tensorrt_engine tensorrt_converter)
@@ -13,3 +13,6 @@ nv_test(test_trt_fc_op SRCS test_fc_op.cc fc_op.cc
         DEPS ${FLUID_CORE_MODULES} tensorrt_engine mul_op SERIAL)
 nv_test(test_trt_activation_op SRCS test_activation_op.cc activation_op.cc
         DEPS ${FLUID_CORE_MODULES} tensorrt_engine activation_op SERIAL)
+
+nv_test(test_trt_pool2d_op SRCS test_pool2d_op.cc pool2d_op.cc
+        DEPS ${FLUID_CORE_MODULES} tensorrt_engine pool_op SERIAL)

paddle/fluid/inference/tensorrt/convert/mul_op.cc

@@ -49,5 +49,4 @@ class MulOpConverter : public OpConverter {
 }  // namespace inference
 }  // namespace paddle
 
-USE_OP(mul);
 REGISTER_TRT_OP_CONVERTER(mul, MulOpConverter);

paddle/fluid/inference/tensorrt/convert/pool2d_op.cc

+/* Copyright (c) 2018 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/fluid/inference/tensorrt/convert/op_converter.h"
+
+namespace paddle {
+namespace inference {
+namespace tensorrt {
+
+/*
+ * Pool2dOp, IPoolingLayer in TRT. This Layer doesn't has weights.
+ */
+class Pool2dOpConverter : public OpConverter {
+ public:
+  void operator()(const framework::proto::OpDesc& op,
+                  const framework::Scope& scope, bool test_mode) override {
+    VLOG(4)
+        << "convert a fluid pool2d op to tensorrt pool2d layer without bias";
+    framework::OpDesc op_desc(op, nullptr);
+    // Declare inputs
+    PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1);
+    PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1);
+    auto* input1 = engine_->GetITensor(op_desc.Input("X")[0]);
+
+    std::string pool_type =
+        boost::get<std::string>(op_desc.GetAttr("pooling_type"));
+    std::vector<int> ksize =
+        boost::get<std::vector<int>>(op_desc.GetAttr("ksize"));
+    std::vector<int> strides =
+        boost::get<std::vector<int>>(op_desc.GetAttr("strides"));
+    std::vector<int> paddings =
+        boost::get<std::vector<int>>(op_desc.GetAttr("paddings"));
+
+    const nvinfer1::DimsHW nv_ksize(ksize[0], ksize[1]);
+    const nvinfer1::DimsHW nv_strides(strides[0], strides[1]);
+    const nvinfer1::DimsHW nv_paddings(paddings[0], paddings[1]);
+
+    PADDLE_ENFORCE_EQ(input1->getDimensions().nbDims, 3UL);
+
+    nvinfer1::PoolingType nv_pool_type = nvinfer1::PoolingType::kMAX;
+    if (pool_type == "max") {
+      nv_pool_type = nvinfer1::PoolingType::kMAX;
+    } else if (pool_type == "avg") {
+      nv_pool_type = nvinfer1::PoolingType::kAVERAGE;
+    } else {
+      PADDLE_THROW("TensorRT unsupported pooling type!");
+    }
+
+    auto* layer = TRT_ENGINE_ADD_LAYER(engine_, Pooling,
+                                       *const_cast<nvinfer1::ITensor*>(input1),
+                                       nv_pool_type, nv_ksize);
+    PADDLE_ENFORCE_NOT_NULL(layer, "pool layer could not be created.");
+    layer->setStride(nv_strides);
+    layer->setPadding(nv_paddings);
+
+    auto output_name = op_desc.Output("Out")[0];
+    engine_->SetITensor(output_name, layer->getOutput(0));
+    if (test_mode) {
+      engine_->DeclareOutput(output_name);
+    }
+  }
+};
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle
+
+USE_OP(pool2d);
+REGISTER_TRT_OP_CONVERTER(pool2d, Pool2dOpConverter);

paddle/fluid/inference/tensorrt/convert/test_activation_op.cc

@@ -37,7 +37,7 @@ TEST(ReluOpConverter, main) {
   validator.SetOp(*desc.Proto());
   LOG(INFO) << "execute";
 
-  validator.Execute(1);
+  validator.Execute(5);
 }
 
 }  // namespace tensorrt

paddle/fluid/inference/tensorrt/convert/test_fc_op.cc

@@ -24,9 +24,8 @@ TEST(fc_op, test) {
   std::unordered_set<std::string> parameters({"mul-Y"});
   framework::Scope scope;
   TRTConvertValidation validator(10, parameters, scope, 1000);
-  validator.DeclInputVar("mul-X", nvinfer1::Dims4(1, 10, 1, 1));
+  validator.DeclInputVar("mul-X", nvinfer1::Dims3(10, 1, 1));
   validator.DeclParamVar("mul-Y", nvinfer1::Dims2(10, 2));
-  // validator.DeclParamVar("mul-Y", nvinfer1::Dims2(8, 2));
   validator.DeclOutputVar("mul-Out", nvinfer1::Dims2(1, 2));
 
   // Prepare Op description
@@ -38,7 +37,7 @@ TEST(fc_op, test) {
 
   validator.SetOp(*desc.Proto());
 
-  validator.Execute(1);
+  validator.Execute(10);
 }
 
 }  // namespace tensorrt

paddle/fluid/inference/tensorrt/convert/test_mul_op.cc

@@ -23,7 +23,7 @@ namespace tensorrt {
 TEST(MulOpConverter, main) {
   framework::Scope scope;
   std::unordered_set<std::string> parameters;
-  TRTConvertValidation validator(10, parameters, scope, 1000);
+  TRTConvertValidation validator(10, parameters, scope, 1000, false);
   validator.DeclInputVar("mul-X", nvinfer1::Dims2(10, 6));
   validator.DeclInputVar("mul-Y", nvinfer1::Dims2(6, 10));
   validator.DeclOutputVar("mul-Out", nvinfer1::Dims2(10, 10));
@@ -39,7 +39,7 @@ TEST(MulOpConverter, main) {
   validator.SetOp(*desc.Proto());
   LOG(INFO) << "execute";
 
-  validator.Execute(1);
+  validator.Execute(2);
 }
 
 }  // namespace tensorrt

paddle/fluid/inference/tensorrt/convert/test_pool2d_op.cc

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+#include <gtest/gtest.h>
+#include <fstream>
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h"
+
+namespace paddle {
+namespace inference {
+namespace tensorrt {
+
+TEST(Pool2dOpConverter, main) {
+  framework::Scope scope;
+  std::unordered_set<std::string> parameters;
+  TRTConvertValidation validator(5, parameters, scope, 1 << 15);
+
+  // The ITensor's Dims should not contain the batch size.
+  // So, the ITensor's Dims of input and output should be C * H * W.
+  validator.DeclInputVar("pool2d-X", nvinfer1::Dims3(3, 4, 4));
+  validator.DeclOutputVar("pool2d-Out", nvinfer1::Dims3(3, 2, 2));
+
+  // Prepare Op description
+  framework::OpDesc desc;
+  desc.SetType("pool2d");
+  desc.SetInput("X", {"pool2d-X"});
+  desc.SetOutput("Out", {"pool2d-Out"});
+
+  std::vector<int> ksize({2, 2});
+  std::vector<int> strides({2, 2});
+  std::vector<int> paddings({0, 0});
+  std::string pooling_t = "max";
+
+  desc.SetAttr("pooling_type", pooling_t);
+  desc.SetAttr("ksize", ksize);
+  desc.SetAttr("strides", strides);
+  desc.SetAttr("paddings", paddings);
+
+  LOG(INFO) << "set OP";
+  validator.SetOp(*desc.Proto());
+  LOG(INFO) << "execute";
+
+  validator.Execute(3);
+}
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle
+
+USE_OP(pool2d);

paddle/fluid/inference/tensorrt/convert/ut_helper.h

@@ -63,13 +63,16 @@ class TRTConvertValidation {
  public:
   TRTConvertValidation() = delete;
 
-  TRTConvertValidation(int batch_size,
+  TRTConvertValidation(int max_batch_size,
                        const std::unordered_set<std::string>& parameters,
                        framework::Scope& scope,  // NOLINT
-                       int workspace_size = 1 << 10)
-      : parameters_(parameters), scope_(scope) {
+                       int workspace_size = 1 << 10, bool if_add_batch = true)
+      : parameters_(parameters),
+        scope_(scope),
+        if_add_batch_(if_add_batch),
+        max_batch_size_(max_batch_size) {
     // create engine.
-    engine_.reset(new TensorRTEngine(batch_size, workspace_size, &stream_));
+    engine_.reset(new TensorRTEngine(max_batch_size, workspace_size, &stream_));
     engine_->InitNetwork();
 
     PADDLE_ENFORCE_EQ(cudaStreamCreate(&stream_), 0);
@@ -84,7 +87,7 @@ class TRTConvertValidation {
 
   // Declare a parameter varaible in the scope.
   void DeclParamVar(const std::string& name, const nvinfer1::Dims& dims) {
-    DeclVar(name, dims);
+    DeclVar(name, dims, true);
   }
 
   void DeclOutputVar(const std::string& name, const nvinfer1::Dims& dims) {
@@ -92,12 +95,18 @@ class TRTConvertValidation {
   }
 
   // Declare a variable in a fluid Scope.
-  void DeclVar(const std::string& name, const nvinfer1::Dims& dims) {
+  void DeclVar(const std::string& name, const nvinfer1::Dims& dims,
+               bool is_param = false) {
     platform::CPUPlace place;
     platform::CPUDeviceContext ctx(place);
 
     // Init Fluid tensor.
     std::vector<int> dim_vec(dims.d, dims.d + dims.nbDims);
+    // There is no batchsize in ITensor's shape, but We should add it to
+    // tensor's shape of fluid. If the variable is not parameter and the
+    // if_add_batch_ flag is true, add the max batchsize to dim_vec.
+    if (is_param != true && if_add_batch_ == true)
+      dim_vec.insert(dim_vec.begin(), max_batch_size_);
     auto* x = scope_.Var(name);
     auto* x_tensor = x->GetMutable<framework::LoDTensor>();
     x_tensor->Resize(framework::make_ddim(dim_vec));
@@ -131,6 +140,7 @@ class TRTConvertValidation {
 
   void Execute(int batch_size) {
     // Execute Fluid Op
+    PADDLE_ENFORCE_LE(batch_size, max_batch_size_);
     platform::CPUPlace place;
     platform::CPUDeviceContext ctx(place);
     op_->Run(scope_, place);
@@ -149,9 +159,15 @@ class TRTConvertValidation {
       auto* var = scope_.FindVar(output);
       auto tensor = var->GetMutable<framework::LoDTensor>();
       framework::TensorToVector(*tensor, ctx, &fluid_out);
+
+      size_t fluid_out_size = fluid_out.size();
+      if (if_add_batch_ == true) {
+        fluid_out_size =
+            batch_size * (framework::product(tensor->dims()) / max_batch_size_);
+      }
       // Compare two output
       ASSERT_FALSE(fluid_out.empty());
-      for (size_t i = 0; i < fluid_out.size(); i++) {
+      for (size_t i = 0; i < fluid_out_size; i++) {
         // Loose the threshold for CI in different machine model.
         EXPECT_LT(std::abs(fluid_out[i] - trt_out[i]), 2e-5);
       }
@@ -167,6 +183,12 @@ class TRTConvertValidation {
   std::unique_ptr<framework::OpDesc> op_desc_;
   const std::unordered_set<std::string>& parameters_;
   framework::Scope& scope_;
+  // The ITensor of trt does not cotain the batch size,
+  // bug, in most cases, we need to set batch size for
+  // fluid's tensor shape. This variable indicates
+  // whether to add batch size to tensor shape of fluid.
+  bool if_add_batch_;
+  int max_batch_size_;
 };
 
 }  // namespace tensorrt

paddle/fluid/inference/tensorrt/test_engine.cc

@@ -113,7 +113,7 @@ TEST_F(TensorRTEngineTest, add_layer_multi_dim) {
   ASSERT_EQ(y_cpu[1], 14.5);
 }
 
-TEST_F(TensorRTEngineTest, test_conv2d_temp) {
+TEST_F(TensorRTEngineTest, test_conv2d) {
   // Weight in CPU memory.
   float raw_weight[9] = {1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0};
   float raw_bias[1] = {0};
@@ -146,6 +146,37 @@ TEST_F(TensorRTEngineTest, test_conv2d_temp) {
   ASSERT_EQ(y_cpu[1], 6.0);
 }
 
+TEST_F(TensorRTEngineTest, test_pool2d) {
+  // Weight in CPU memory.
+  auto* x = engine_->DeclareInput("x", nvinfer1::DataType::kFLOAT,
+                                  nvinfer1::Dims3{1, 2, 2});
+
+  nvinfer1::PoolingType pool_t = nvinfer1::PoolingType::kAVERAGE;
+  auto* pool_layer =
+      TRT_ENGINE_ADD_LAYER(engine_, Pooling, *const_cast<nvinfer1::ITensor*>(x),
+                           pool_t, nvinfer1::DimsHW{2, 2});
+
+  PADDLE_ENFORCE(pool_layer != nullptr);
+  pool_layer->setStride(nvinfer1::DimsHW{1, 1});
+  pool_layer->setPadding(nvinfer1::DimsHW{0, 0});
+
+  engine_->DeclareOutput(pool_layer, 0, "y");
+  engine_->FreezeNetwork();
+  ASSERT_EQ(engine_->engine()->getNbBindings(), 2);
+
+  float x_v[8] = {1.0, 2.0, 5.0, 0.0, 2.0, 3.0, 5.0, 10.0};
+  engine_->SetInputFromCPU("x", reinterpret_cast<void*>(&x_v),
+                           8 * sizeof(float));
+  engine_->Execute(2);
+
+  LOG(INFO) << "to get output";
+  float* y_cpu = new float[2];
+  engine_->GetOutputInCPU("y", &y_cpu[0], 2 * sizeof(float));
+
+  ASSERT_EQ(y_cpu[0], 2.0);
+  ASSERT_EQ(y_cpu[1], 5.0);
+}
+
 }  // namespace tensorrt
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/tests/book/CMakeLists.txt

@@ -17,7 +17,7 @@ function(inference_test TARGET_NAME)
     string(REGEX REPLACE "^_$" "" arg "${arg}")
     cc_test(test_inference_${TARGET_NAME}${arg}
         SRCS test_inference_${TARGET_NAME}.cc
-        DEPS paddle_fluid
+        DEPS paddle_fluid_origin
         ARGS --dirname=${PYTHON_TESTS_DIR}/book/${TARGET_NAME}${arg}.inference.model)
     set_tests_properties(test_inference_${TARGET_NAME}${arg}
         PROPERTIES DEPENDS test_${TARGET_NAME})
@@ -43,6 +43,6 @@ inference_test(word2vec)
 # TODO(TJ): clean me up
 cc_test(test_inference_nlp
   SRCS test_inference_nlp.cc
-  DEPS paddle_fluid
+  DEPS paddle_fluid_origin
   ARGS
   --model_path=${PADDLE_BINARY_DIR}/python/paddle/fluid/tests/book/recognize_digits_mlp.inference.model)

paddle/fluid/inference/tests/book/test_inference_nlp.cc

@@ -20,9 +20,6 @@ limitations under the License. */
 #include "gtest/gtest.h"
 #include "paddle/fluid/inference/tests/test_helper.h"
 #include "paddle/fluid/platform/cpu_helper.h"
-#ifdef PADDLE_WITH_MKLML
-#include <omp.h>
-#endif
 
 DEFINE_string(model_path, "", "Directory of the inference model.");
 DEFINE_string(data_file, "", "File of input index data.");
@@ -30,6 +27,7 @@ DEFINE_int32(repeat, 100, "Running the inference program repeat times");
 DEFINE_bool(prepare_vars, true, "Prepare variables before executor");
 DEFINE_int32(num_threads, 1, "Number of threads should be used");
 DECLARE_bool(use_mkldnn);
+DECLARE_int32(paddle_num_threads);
 
 inline double GetCurrentMs() {
   struct timeval time;
@@ -160,12 +158,7 @@ TEST(inference, nlp) {
   std::unique_ptr<paddle::framework::Scope> scope(
       new paddle::framework::Scope());
 
-#ifdef PADDLE_WITH_MKLML
-  // only use 1 thread number per std::thread
-  omp_set_dynamic(0);
-  omp_set_num_threads(1);
-  paddle::platform::SetNumThreads(1);
-#endif
+  paddle::platform::SetNumThreads(FLAGS_paddle_num_threads);
 
   double start_ms = 0, stop_ms = 0;
   if (FLAGS_num_threads > 1) {

paddle/fluid/operators/CMakeLists.txt

@@ -270,6 +270,9 @@ op_library(cos_sim_op DEPS cos_sim_functor)
 op_library(parallel_do_op DEPS executor)
 op_library(unsqueeze_op DEPS reshape_op)
 op_library(squeeze_op DEPS reshape_op)
+op_library(extract_rows_op DEPS memory)
+op_library(flatten_op DEPS reshape_op)
+
 
 if (WITH_GPU)
     op_library(conv_op DEPS vol2col depthwise_conv im2col)

paddle/fluid/operators/distributed/CMakeLists.txt

@@ -17,9 +17,9 @@ if(WITH_GRPC)
   set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
   set_source_files_properties(grpc_serde_test.cc rpc_server_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
   cc_test(grpc_serde_test SRCS grpc_serde_test.cc 
-     DEPS grpc++_unsecure grpc_unsecure gpr cares zlib protobuf sendrecvop_grpc scope profiler math_function SERIAL)
-  cc_test(rpc_server_test SRCS rpc_server_test.cc 
-    DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf executor  proto_desc lookup_table_op SERIAL)
+    DEPS grpc++_unsecure grpc_unsecure gpr cares zlib protobuf sendrecvop_grpc scope profiler math_function SERIAL)
+  cc_test(rpc_server_test SRCS rpc_server_test.cc
+    DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf executor  proto_desc lookup_sparse_table_op SERIAL)
   return()
 endif()
 

paddle/fluid/operators/distributed/grpc_client.cc

@@ -49,6 +49,7 @@ void GRPCClient::SendComplete() {
 }
 
 GRPCClient::~GRPCClient() {
+  stopped_ = true;
   Wait();
   cq_.Shutdown();
   {
@@ -275,7 +276,7 @@ void GRPCClient::Proceed() {
   void* tag = nullptr;
   bool ok = false;
 
-  while (cq_.Next(&tag, &ok)) {
+  while (!stopped_ && cq_.Next(&tag, &ok)) {
     BaseProcessor* c = static_cast<BaseProcessor*>(tag);
     GPR_ASSERT(ok);
     PADDLE_ENFORCE(c);

paddle/fluid/operators/distributed/grpc_client.h

@@ -174,7 +174,7 @@ class CheckpointNotifyProcessor : public BaseProcessor {
 
 class GRPCClient : public RPCClient {
  public:
-  GRPCClient() : ok_(true), completed_(false) {}
+  GRPCClient() : ok_(true), completed_(false), stopped_(false) {}
   virtual ~GRPCClient();
 
   bool AsyncSendVar(const std::string& ep, const platform::DeviceContext& ctx,
@@ -237,6 +237,8 @@ class GRPCClient : public RPCClient {
   // mutex for sending complete message only once
   std::mutex completed_mutex_;
   bool completed_;
+
+  volatile bool stopped_;
 };
 
 }  // namespace distributed

paddle/fluid/operators/distributed/rpc_server_test.cc

@@ -30,7 +30,7 @@ namespace framework = paddle::framework;
 namespace platform = paddle::platform;
 namespace distributed = paddle::operators::distributed;
 
-USE_OP(lookup_table);
+USE_NO_KERNEL_OP(lookup_sparse_table);
 
 std::unique_ptr<distributed::RPCServer> g_rpc_service;
 std::unique_ptr<distributed::RequestHandler> g_req_handler;
@@ -42,13 +42,13 @@ framework::BlockDesc* AppendPrefetchBlcok(framework::ProgramDesc* program) {
   framework::VariableNameMap input({{"W", {"w"}}, {"Ids", {"ids"}}});
   framework::VariableNameMap output({{"Output", {"out"}}});
   auto op = block->AppendOp();
-  op->SetType("lookup_table");
+  op->SetType("lookup_sparse_table");
   op->SetInput("W", {"w"});
   op->SetInput("Ids", {"ids"});
   op->SetOutput("Out", {"out"});
 
   auto& out = *root_block->Var("out");
-  out.SetType(framework::proto::VarType::SELECTED_ROWS);
+  out.SetType(framework::proto::VarType::LOD_TENSOR);
   out.SetShape({10, 10});
 
   return block;
@@ -59,20 +59,19 @@ void CreateVarsOnScope(framework::Scope* scope, platform::CPUPlace* place) {
   w_var->GetMutable<framework::SelectedRows>();
 
   auto out_var = scope->Var("out");
-  out_var->GetMutable<framework::SelectedRows>();
+  out_var->GetMutable<framework::LoDTensor>();
 
   auto ids_var = scope->Var("ids");
-  ids_var->GetMutable<framework::SelectedRows>();
+  ids_var->GetMutable<framework::LoDTensor>();
 }
 
 void InitTensorsOnClient(framework::Scope* scope, platform::CPUPlace* place,
                          int64_t rows_numel) {
   CreateVarsOnScope(scope, place);
-  auto ids_var = scope->Var("ids")->GetMutable<framework::SelectedRows>();
-  auto rows = ids_var->mutable_rows();
-  for (int64_t i = 0; i < rows_numel; ++i) rows->push_back(i * 2);
-  ids_var->mutable_value()->Resize({rows_numel, 1});
-  ids_var->mutable_value()->mutable_data<float>(*place);
+  auto ids_var = scope->Var("ids")->GetMutable<framework::LoDTensor>();
+  int64_t* ids_ptr =
+      ids_var->mutable_data<int64_t>(framework::DDim({rows_numel, 1}), *place);
+  for (int64_t i = 0; i < rows_numel; ++i) ids_ptr[i] = i * 2;
 }
 
 void InitTensorsOnServer(framework::Scope* scope, platform::CPUPlace* place,
@@ -148,11 +147,11 @@ TEST(PREFETCH, CPU) {
     client->AsyncPrefetchVar(ep, ctx, scope, in_var_name, out_var_name);
     client->Wait();
     auto var = scope.Var(out_var_name);
-    auto value = var->GetMutable<framework::SelectedRows>()->value();
-    auto ptr = value.mutable_data<float>(place);
+    auto value = var->GetMutable<framework::LoDTensor>();
+    auto ptr = value->mutable_data<float>(place);
 
     for (int64_t i = 0; i < rows_numel; ++i) {
-      EXPECT_EQ(ptr[0 + i * value.dims()[1]], static_cast<float>(i * 2));
+      EXPECT_EQ(ptr[0 + i * value->dims()[1]], static_cast<float>(i * 2));
     }
   }
 

paddle/fluid/operators/extract_rows_op.cc

+/* Copyright (c) 2018 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 <string>
+#include <vector>
+#include "paddle/fluid/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+class ExtractRowsOpInferShape : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of ExtractRowsOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of ExtractRowsOp should not be null.");
+    PADDLE_ENFORCE_EQ(ctx->GetInputsVarType("X")[0],
+                      framework::proto::VarType::SELECTED_ROWS,
+                      "The type of input(X) must be SelectedRows.");
+    auto in_dims = ctx->GetInputDim("X");
+
+    ctx->SetOutputDim(
+        "Out", framework::make_ddim(std::vector<int64_t>{in_dims[0], 1}));
+  }
+};
+
+class ExtractRowsOp : public framework::OperatorBase {
+ public:
+  ExtractRowsOp(const std::string &type,
+                const framework::VariableNameMap &inputs,
+                const framework::VariableNameMap &outputs,
+                const framework::AttributeMap &attrs)
+      : framework::OperatorBase(type, inputs, outputs, attrs) {}
+
+ private:
+  void RunImpl(const framework::Scope &scope,
+               const platform::Place &place) const override {
+    auto &in = scope.FindVar(Input("X"))->Get<framework::SelectedRows>();
+    auto out = scope.FindVar(Output("Out"))->GetMutable<framework::LoDTensor>();
+
+    auto in_rows = in.rows();
+    auto out_dim = framework::make_ddim(
+        std::vector<int64_t>{static_cast<int64_t>(in_rows.size()), 1});
+    auto dst_ptr = out->mutable_data<int64_t>(out_dim, in.place());
+
+    if (paddle::platform::is_gpu_place(in.place())) {
+#ifdef PADDLE_WITH_CUDA
+      platform::DeviceContextPool &pool =
+          platform::DeviceContextPool::Instance();
+      auto *dev_ctx = pool.Get(in.place());
+      auto src_ptr = in_rows.Data(in.place());
+      auto stream =
+          reinterpret_cast<const platform::CUDADeviceContext &>(*dev_ctx)
+              .stream();
+      memory::Copy(boost::get<platform::CUDAPlace>(out->place()), dst_ptr,
+                   boost::get<platform::CUDAPlace>(in.place()), src_ptr,
+                   in_rows.size() * sizeof(int64_t), stream);
+#else
+      PADDLE_THROW("Not compiled with CUDA.");
+#endif
+    } else {
+      memory::Copy(platform::CPUPlace(), dst_ptr, platform::CPUPlace(),
+                   in_rows.data(), in_rows.size() * sizeof(int64_t));
+    }
+  }
+};
+
+class ExtractRowsOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X",
+             "(SelectedRows). The input tensor of extract_rows operator,"
+             " and its type is SelectedRows.");
+    AddOutput("Out", "(Tensor). The the rows of input(X).");
+
+    AddComment(R"DOC(
+    ExtractRows Operator.
+
+The function of extract_rows_op is extracting the rows from the input(X)
+whose type is SelectedRows.
+
+    )DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(extract_rows, ops::ExtractRowsOp, ops::ExtractRowsOpMaker,
+                  ops::ExtractRowsOpInferShape);

paddle/fluid/operators/flatten_op.cc

+/* Copyright (c) 2018 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 <vector>
+#include "paddle/fluid/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+class FlattenOpInferShape : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input (X) of Flatten op should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output (Output) of Flatten op should not be null.");
+    const auto &axis = ctx->Attrs().Get<int>("axis");
+    const auto &in_dims = ctx->GetInputDim("X");
+    PADDLE_ENFORCE(axis >= 0, "The axis should be greater than or equal to 0.");
+    PADDLE_ENFORCE(
+        axis <= in_dims.size(),
+        "The axis should be less than or equal to input tensor's rank.");
+
+    const auto &out_dims = GetOutputShape(axis, in_dims);
+    ctx->SetOutputDim("Out", framework::make_ddim(out_dims));
+    if (in_dims[0] == out_dims[0]) {
+      // Only pass LoD when the first dimension of output and Input(X)
+      // are the same.
+      ctx->ShareLoD("X", "Out");
+    }
+  }
+
+  static std::vector<int32_t> GetOutputShape(const int axis,
+                                             const framework::DDim &in_dims) {
+    int64_t outer = 1, inner = 1;
+    for (int i = 0; i < in_dims.size(); ++i) {
+      if (i < axis) {
+        outer *= in_dims[i];
+      } else {
+        inner *= in_dims[i];
+      }
+    }
+    std::vector<int32_t> out_shape(2);
+    out_shape[0] = outer;
+    out_shape[1] = inner;
+    return out_shape;
+  }
+};
+
+class FlattenOp : public framework::OperatorBase {
+ public:
+  using OperatorBase::OperatorBase;
+
+ private:
+  void RunImpl(const framework::Scope &scope,
+               const platform::Place &place) const override {
+    auto &axis = Attr<int>("axis");
+    auto in_dims =
+        scope.FindVar(Input("X"))->Get<framework::LoDTensor>().dims();
+    const auto &out_dims = FlattenOpInferShape::GetOutputShape(axis, in_dims);
+
+    framework::AttributeMap attrs;
+    attrs["shape"] = out_dims;
+    attrs["inplace"] = false;
+    // Invoke Reshape Op
+    auto reshape_op = framework::OpRegistry::CreateOp(
+        "reshape", {{"X", {Input("X")}}, {"Shape", {}}},
+        {{"Out", {Output("Out")}}}, attrs);
+    reshape_op->Run(scope, place);
+  }
+};
+
+class FlattenOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X", "(Tensor) A tensor of rank >= axis.");
+    AddOutput("Out",
+              "A 2D tensor is reshaped input tensor. The input dimensions"
+              "up to axis are flattened to the outer dimension of the output"
+              "and the remaining input dimensions are flattened into the inner"
+              "dimension of the output.");
+    AddAttr<int>("axis",
+                 "(int)"
+                 "Indicate up to which input dimensions (exclusive) should be"
+                 "flattened to the outer dimension of the output. The value"
+                 "for axis must be in the range [0, R], where R is the rank of"
+                 "the input tensor. When axis = 0, the shape of the output"
+                 "tensor is (1, (d_0 X d_1 ... d_n), where the shape of the"
+                 "input tensor is (d_0, d_1, ... d_n).")
+        .SetDefault(1);
+    AddComment(R"DOC(
+Flatten Operator
+
+Flattens the input tensor into a 2D matrix.
+
+Examples:
+Case 1:
+  Given
+    X.shape = (3, 100, 100, 4)
+  and
+    axis = 2
+  We get:
+    Out.shape = (3 * 100, 4 * 100)
+
+Case 2:
+  Given
+    X.shape = (3, 100, 100, 4)
+  and
+    axis = 0
+  We get:
+    Out.shape = (1, 3 * 100 * 100 * 4)
+)DOC");
+  }
+};
+
+class FlattenGradInferShape : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext *context) const override {
+    context->SetOutputDim(framework::GradVarName("X"),
+                          context->GetInputDim("X"));
+    context->ShareLoD("X", framework::GradVarName("X"));
+  }
+};
+
+class FlattenGradOp : public framework::OperatorBase {
+ public:
+  using OperatorBase::OperatorBase;
+
+ private:
+  void RunImpl(const framework::Scope &scope,
+               const platform::Place &place) const override {
+    auto dx_name = Output(framework::GradVarName("X"));
+    auto dout_name = Input(framework::GradVarName("Out"));
+    auto in_dims =
+        scope.FindVar(Input("X"))->Get<framework::LoDTensor>().dims();
+    framework::AttributeMap attrs;
+    attrs["shape"] = framework::vectorize2int(in_dims);
+    attrs["inplace"] = false;
+
+    auto reshape_op = framework::OpRegistry::CreateOp(
+        "reshape", {{"X", {dout_name}}, {"Shape", {}}}, {{"Out", {dx_name}}},
+        attrs);
+    reshape_op->Run(scope, place);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+USE_OP(reshape);
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(flatten, ops::FlattenOp, ops::FlattenOpMaker,
+                  ops::FlattenOpInferShape,
+                  paddle::framework::DefaultGradOpDescMaker<true>);
+REGISTER_OPERATOR(flatten_grad, ops::FlattenGradOp, ops::FlattenGradInferShape);

paddle/fluid/operators/lookup_table_op.cc

@@ -33,19 +33,15 @@ class LookupTableOp : public framework::OperatorWithKernel {
     auto table_dims = ctx->GetInputDim("W");
     auto ids_dims = ctx->GetInputDim("Ids");
 
-    auto ids_var_type = ctx->GetInputsVarType("Ids").front();
-    // The type of Ids(Input) is SelectedRows or LoDTensor, when Ids's type
-    // is LoDTensor, this tensor contains the ids to be looked up in W
-    // and it must be a column vector with rank = 2 while the 2nd dimension
-    // size must be 1, when Ids's type is SelectedRows, the rows of Ids
-    // contains the ids to be looked up in W;
-    if (ids_var_type == framework::proto::VarType::LOD_TENSOR) {
-      PADDLE_ENFORCE_EQ(ids_dims.size(), 2);
-      PADDLE_ENFORCE_EQ(ids_dims[1], 1);
-    }
+    PADDLE_ENFORCE_EQ(ids_dims.size(), 2);
+    PADDLE_ENFORCE_EQ(ids_dims[1], 1);
 
     ctx->SetOutputDim("Out", {ids_dims[0], table_dims[1]});
-    ctx->ShareLoD("Ids", /*->*/ "Out");
+
+    if (ctx->GetOutputsVarType("Out")[0] ==
+        framework::proto::VarType::LOD_TENSOR) {
+      ctx->ShareLoD("Ids", /*->*/ "Out");
+    }
   }
 
  protected:
@@ -62,17 +58,12 @@ class LookupTableOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("W",
              "(Tensor) The input represents embedding tensors, "
              "which is a learnable parameter.");
-    AddInput(
-        "Ids",
-        "(Tensor or SelectedRows) Ids's type can be Tensor or "
-        "SelectedRows, when Ids's type is Tensor, this tensor contains "
-        "the ids to be looked up in W and it must be a column vector with "
-        "rank = 2 while the 2nd dimension size must be 1; when Ids's type is "
-        "SelectedRows, the rows of Ids contains the ids to be looked up "
-        "in W.");
-    AddOutput("Out",
-              "(Tensor or SelectedRows) The lookup results, which have the "
-              "same type as W.");
+    AddInput("Ids",
+             "An input with type int32 or int64 "
+             "contains the ids to be looked up in W. "
+             "Ids must be a column vector with rank = 2. "
+             "The 2nd dimension size must be 1.");
+    AddOutput("Out", "The lookup results, which have the same type as W.");
     AddAttr<bool>("is_sparse",
                   "(boolean, default false) "
                   "Sparse update.")
@@ -90,15 +81,10 @@ class LookupTableOpMaker : public framework::OpProtoAndCheckerMaker {
 Lookup Table Operator.
 
 This operator is used to perform lookups on the parameter W,
-then concatenated into a dense or sparse tensor.
-
-The type of Ids(Input) is SelectedRows, Tensor or LoDTensor, when Ids's
-type is SelectedRows, the rows of Ids contains the ids to be looked up in W;
-when Ids's type is Tensor, this tensor contains the ids to be looked up in W
-and it must be a column vector with rank = 2 while the 2nd dimension size must be 1,
-at this time, Ids can carry the LoD (Level of Details) information, or not, and
-the output only shares the LoD information with input Ids.
+then concatenated into a dense tensor.
 
+The input Ids can carry the LoD (Level of Details) information,
+or not. And the output only shares the LoD information with input Ids.
 
 )DOC");
   }

paddle/fluid/operators/lookup_table_op.cu

@@ -23,7 +23,7 @@ namespace operators {
 
 template <typename T, int BlockDimX, int BlockDimY, int GridDimX,
           bool PaddingFlag>
-__global__ void LookupTable(T* output, const T* table, const int64_t* ids,
+__global__ void LookupTable(T *output, const T *table, const int64_t *ids,
                             const int64_t N, const int64_t K, const int64_t D,
                             const int64_t padding_idx) {
   int idx = threadIdx.x;
@@ -33,8 +33,8 @@ __global__ void LookupTable(T* output, const T* table, const int64_t* ids,
     int64_t id = ids[idy];
     PADDLE_ASSERT(id >= 0);
     PADDLE_ASSERT(id < N);
-    T* out = output + idy * D;
-    const T* tab = table + id * D;
+    T *out = output + idy * D;
+    const T *tab = table + id * D;
     for (int i = idx; i < D; i += BlockDimX) {
       if (PaddingFlag) {
         if (id == padding_idx)
@@ -50,7 +50,7 @@ __global__ void LookupTable(T* output, const T* table, const int64_t* ids,
 }
 
 template <typename T, int BlockDimX, int BlockDimY, int GridDimX>
-__global__ void LookupTableGrad(T* table, const T* output, const int64_t* ids,
+__global__ void LookupTableGrad(T *table, const T *output, const int64_t *ids,
                                 const int64_t N, const int64_t K,
                                 const int64_t D) {
   int idx = threadIdx.x;
@@ -60,8 +60,8 @@ __global__ void LookupTableGrad(T* table, const T* output, const int64_t* ids,
     int id = ids[idy];
     PADDLE_ASSERT(id >= 0);
     PADDLE_ASSERT(id < N);
-    const T* out = output + idy * D;
-    T* tab = table + id * D;
+    const T *out = output + idy * D;
+    T *tab = table + id * D;
     for (int i = idx; i < D; i += BlockDimX) {
       paddle::platform::CudaAtomicAdd(&tab[i], out[i]);
     }
@@ -72,36 +72,19 @@ __global__ void LookupTableGrad(T* table, const T* output, const int64_t* ids,
 template <typename T>
 class LookupTableCUDAKernel : public framework::OpKernel<T> {
  public:
-  void Compute(const framework::ExecutionContext& context) const override {
-    auto* table_t = context.Input<LoDTensor>("W");
+  void Compute(const framework::ExecutionContext &context) const override {
+    auto *table_t = context.Input<LoDTensor>("W");
+    auto *ids_t = context.Input<LoDTensor>("Ids");
+    auto *output_t = context.Output<LoDTensor>("Out");
     int64_t padding_idx = context.Attr<int64_t>("padding_idx");
-    auto* ids_var = context.InputVar("Ids");
-    Tensor* output_t = context.Output<Tensor>("Out");
-
-    int64_t* ids;
-    int64_t K;
-
-    // The type of Ids(Input) is SelectedRows or LoDTensor, when Ids's type
-    // is LoDTensor, this tensor contains the ids to be looked up in W;
-    // when Ids's type is SelectedRows, the rows of Ids contains the
-    // ids to be looked up in W.
-    if (ids_var->IsType<framework::LoDTensor>()) {
-      auto* ids_t = context.Input<LoDTensor>("Ids");
-      ids = const_cast<int64_t*>(ids_t->data<int64_t>());
-      K = ids_t->numel();
-    } else if (ids_var->IsType<framework::SelectedRows>()) {
-      auto* ids_t = context.Input<framework::SelectedRows>("Ids");
-      ids = const_cast<int64_t*>(ids_t->rows().CUDAData(context.GetPlace()));
-      K = ids_t->rows().size();
-      output_t->Resize({K, table_t->dims()[1]});
-    } else {
-      PADDLE_THROW("Unsupported Variable Type of Ids");
-    }
 
     size_t N = table_t->dims()[0];
     size_t D = table_t->dims()[1];
-    auto* table = table_t->data<T>();
-    auto* output = output_t->mutable_data<T>(context.GetPlace());
+    size_t K = ids_t->numel();
+
+    auto *ids = ids_t->data<int64_t>();
+    auto *table = table_t->data<T>();
+    auto *output = output_t->mutable_data<T>(context.GetPlace());
 
     dim3 threads(128, 8);
     dim3 grids(8, 1);
@@ -122,19 +105,19 @@ class LookupTableCUDAKernel : public framework::OpKernel<T> {
 template <typename T>
 class LookupTableGradCUDAKernel : public framework::OpKernel<T> {
  public:
-  void Compute(const framework::ExecutionContext& context) const override {
-    auto& dev_ctx =
+  void Compute(const framework::ExecutionContext &context) const override {
+    auto &dev_ctx =
         context.template device_context<platform::CUDADeviceContext>();
     bool is_sparse = context.Attr<bool>("is_sparse");
     // Since paddings are not trainable and fixed in forward, the gradient of
     // paddings makes no sense and we don't deal with it in backward.
     if (is_sparse) {
-      auto* ids = context.Input<LoDTensor>("Ids");
-      auto* table = context.Input<LoDTensor>("W");
-      auto* d_output = context.Input<LoDTensor>(framework::GradVarName("Out"));
-      auto* d_table = context.Output<SelectedRows>(framework::GradVarName("W"));
+      auto *ids = context.Input<LoDTensor>("Ids");
+      auto *table = context.Input<LoDTensor>("W");
+      auto *d_output = context.Input<LoDTensor>(framework::GradVarName("Out"));
+      auto *d_table = context.Output<SelectedRows>(framework::GradVarName("W"));
 
-      auto* ids_data = ids->data<int64_t>();
+      auto *ids_data = ids->data<int64_t>();
       auto ids_dim = ids->dims();
 
       auto stream = dev_ctx.stream();
@@ -150,12 +133,12 @@ class LookupTableGradCUDAKernel : public framework::OpKernel<T> {
 
       d_table->set_rows(new_rows);
 
-      auto* d_table_value = d_table->mutable_value();
+      auto *d_table_value = d_table->mutable_value();
       d_table_value->Resize({ids_dim[0], table->dims()[1]});
       d_table_value->mutable_data<T>(context.GetPlace());
 
-      auto* d_table_data = d_table_value->data<T>();
-      auto* d_output_data = d_output->data<T>();
+      auto *d_table_data = d_table_value->data<T>();
+      auto *d_output_data = d_output->data<T>();
       PADDLE_ENFORCE_EQ(d_table_value->dims(), d_output->dims());
       memory::Copy(gpu_place, d_table_data, gpu_place, d_output_data,
                    d_output->numel() * sizeof(T), stream);
@@ -168,9 +151,9 @@ class LookupTableGradCUDAKernel : public framework::OpKernel<T> {
       int N = d_table_t->dims()[0];
       int D = d_table_t->dims()[1];
       int K = ids_t->numel();
-      const int64_t* ids = ids_t->data<int64_t>();
-      const T* d_output = d_output_t->data<T>();
-      T* d_table = d_table_t->mutable_data<T>(context.GetPlace());
+      const int64_t *ids = ids_t->data<int64_t>();
+      const T *d_output = d_output_t->data<T>();
+      T *d_table = d_table_t->mutable_data<T>(context.GetPlace());
 
       auto t = framework::EigenVector<T>::Flatten(*d_table_t);
       t.device(*dev_ctx.eigen_device()) = t.constant(static_cast<T>(0));

paddle/fluid/operators/lookup_table_op.h

@@ -36,43 +36,13 @@ template <typename T>
 class LookupTableKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &context) const override {
+    auto *ids_t = context.Input<LoDTensor>("Ids");      // int tensor
+    auto *output_t = context.Output<LoDTensor>("Out");  // float tensor
     auto *table_var = context.InputVar("W");
-    auto *ids_var = context.InputVar("Ids");
-    Tensor *output_t = context.Output<Tensor>("Out");
-    int64_t padding_idx = context.Attr<int64_t>("padding_idx");
-
-    DDim table_dim;
 
-    if (table_var->IsType<LoDTensor>()) {
-      table_dim = context.Input<LoDTensor>("W")->dims();
-    } else if (table_var->IsType<SelectedRows>()) {
-      auto *table_t = context.Input<SelectedRows>("W");
-      table_dim = table_t->value().dims();
-    } else {
-      PADDLE_THROW(
-          "The parameter W of a LookupTable "
-          "must be either LoDTensor or SelectedRows");
-    }
-
-    int64_t *ids;
-    int64_t ids_numel;
-
-    // The type of Ids(Input) is SelectedRows or LoDTensor, when Ids's type
-    // is LoDTensor, this tensor contains the ids to be looked up in W;
-    // when Ids's type is SelectedRows, the rows of Ids contains the
-    // ids to be looked up in W.
-    if (ids_var->IsType<LoDTensor>()) {
-      auto *ids_t = context.Input<LoDTensor>("Ids");
-      ids = const_cast<int64_t *>(ids_t->data<int64_t>());
-      ids_numel = ids_t->numel();
-    } else if (ids_var->IsType<SelectedRows>()) {
-      auto *ids_t = context.Input<SelectedRows>("Ids");
-      ids = const_cast<int64_t *>(ids_t->rows().data());
-      ids_numel = ids_t->rows().size();
-      output_t->Resize({ids_numel, table_dim[1]});
-    } else {
-      PADDLE_THROW("Unsupported Variable Type of Ids");
-    }
+    int64_t padding_idx = context.Attr<int64_t>("padding_idx");
+    int64_t *ids = const_cast<int64_t *>(ids_t->data<int64_t>());
+    int64_t ids_numel = ids_t->numel();
 
     if (table_var->IsType<LoDTensor>()) {
       auto *table_t = context.Input<LoDTensor>("W");

paddle/fluid/operators/tensorrt_engine_op.cc

@@ -163,7 +163,4 @@ REGISTER_OP_CPU_KERNEL(
     ops::TensorRTEngineKernel<paddle::platform::CPUDeviceContext, int>,
     ops::TensorRTEngineKernel<paddle::platform::CPUDeviceContext, int64_t>);
 
-// A trick to compile with the needed TensorRT op converter.
-USE_TRT_CONVERTER(mul)
-
 #endif  // PADDLE_WITH_CUDA

paddle/fluid/platform/cpu_helper.cc

@@ -16,6 +16,7 @@ limitations under the License. */
 #include "paddle/fluid/platform/enforce.h"
 
 #ifdef PADDLE_WITH_MKLML
+#include <omp.h>
 #include "paddle/fluid/platform/dynload/mklml.h"
 #endif
 
@@ -33,6 +34,7 @@ void SetNumThreads(int num_threads) {
 #elif defined(PADDLE_WITH_MKLML)
   int real_num_threads = num_threads > 1 ? num_threads : 1;
   platform::dynload::MKL_Set_Num_Threads(real_num_threads);
+  omp_set_num_threads(num_threads);
 #else
   PADDLE_ENFORCE(false, "To be implemented.");
 #endif

paddle/fluid/platform/init.cc

@@ -23,6 +23,9 @@ limitations under the License. */
 #include "paddle/fluid/platform/place.h"
 #include "paddle/fluid/string/piece.h"
 
+DEFINE_int32(paddle_num_threads, 1,
+             "Number of threads for each paddle instance.");
+
 namespace paddle {
 namespace framework {
 
@@ -115,7 +118,7 @@ void InitDevices(bool init_p2p, const std::vector<int> devices) {
   places.emplace_back(platform::CPUPlace());
   platform::DeviceContextPool::Init(places);
 #ifndef PADDLE_WITH_MKLDNN
-  platform::SetNumThreads(1);
+  platform::SetNumThreads(FLAGS_paddle_num_threads);
 #endif
 }
 

patches/grpc/completion_queue.h

+// Copyright (c) 2018 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.
+
+/// A completion queue implements a concurrent producer-consumer queue, with
+/// two main API-exposed methods: \a Next and \a AsyncNext. These
+/// methods are the essential component of the gRPC C++ asynchronous API.
+/// There is also a \a Shutdown method to indicate that a given completion queue
+/// will no longer have regular events. This must be called before the
+/// completion queue is destroyed.
+/// All completion queue APIs are thread-safe and may be used concurrently with
+/// any other completion queue API invocation; it is acceptable to have
+/// multiple threads calling \a Next or \a AsyncNext on the same or different
+/// completion queues, or to call these methods concurrently with a \a Shutdown
+/// elsewhere.
+/// \remark{All other API calls on completion queue should be completed before
+/// a completion queue destructor is called.}
+#ifndef GRPCPP_IMPL_CODEGEN_COMPLETION_QUEUE_H
+#define GRPCPP_IMPL_CODEGEN_COMPLETION_QUEUE_H
+
+#include <typeinfo>
+
+#include <grpc/impl/codegen/atm.h>
+#include <grpcpp/impl/codegen/completion_queue_tag.h>
+#include <grpcpp/impl/codegen/core_codegen_interface.h>
+#include <grpcpp/impl/codegen/grpc_library.h>
+#include <grpcpp/impl/codegen/status.h>
+#include <grpcpp/impl/codegen/time.h>
+
+struct grpc_completion_queue;
+
+namespace grpc {
+
+template <class R>
+class ClientReader;
+template <class W>
+class ClientWriter;
+template <class W, class R>
+class ClientReaderWriter;
+template <class R>
+class ServerReader;
+template <class W>
+class ServerWriter;
+namespace internal {
+template <class W, class R>
+class ServerReaderWriterBody;
+}  // namespace internal
+
+class Channel;
+class ChannelInterface;
+class ClientContext;
+class CompletionQueue;
+class Server;
+class ServerBuilder;
+class ServerContext;
+class ServerInterface;
+
+namespace internal {
+class CompletionQueueTag;
+class RpcMethod;
+template <class ServiceType, class RequestType, class ResponseType>
+class RpcMethodHandler;
+template <class ServiceType, class RequestType, class ResponseType>
+class ClientStreamingHandler;
+template <class ServiceType, class RequestType, class ResponseType>
+class ServerStreamingHandler;
+template <class ServiceType, class RequestType, class ResponseType>
+class BidiStreamingHandler;
+class UnknownMethodHandler;
+template <class Streamer, bool WriteNeeded>
+class TemplatedBidiStreamingHandler;
+template <class InputMessage, class OutputMessage>
+class BlockingUnaryCallImpl;
+}  // namespace internal
+
+extern CoreCodegenInterface* g_core_codegen_interface;
+
+/// A thin wrapper around \ref grpc_completion_queue (see \ref
+/// src/core/lib/surface/completion_queue.h).
+/// See \ref doc/cpp/perf_notes.md for notes on best practices for high
+/// performance servers.
+class CompletionQueue : private GrpcLibraryCodegen {
+ public:
+  /// Default constructor. Implicitly creates a \a grpc_completion_queue
+  /// instance.
+  CompletionQueue()
+      : CompletionQueue(grpc_completion_queue_attributes{
+            GRPC_CQ_CURRENT_VERSION, GRPC_CQ_NEXT, GRPC_CQ_DEFAULT_POLLING}) {}
+
+  /// Wrap \a take, taking ownership of the instance.
+  ///
+  /// \param take The completion queue instance to wrap. Ownership is taken.
+  explicit CompletionQueue(grpc_completion_queue* take);
+
+  /// Destructor. Destroys the owned wrapped completion queue / instance.
+  ~CompletionQueue() {
+    if (typeid(*g_core_codegen_interface).hash_code() !=
+        typeid(CoreCodegenInterface).hash_code()) {
+      g_core_codegen_interface->grpc_completion_queue_destroy(cq_);
+    }
+  }
+
+  /// Tri-state return for AsyncNext: SHUTDOWN, GOT_EVENT, TIMEOUT.
+  enum NextStatus {
+    SHUTDOWN,   ///< The completion queue has been shutdown and fully-drained
+    GOT_EVENT,  ///< Got a new event; \a tag will be filled in with its
+                ///< associated value; \a ok indicating its success.
+    TIMEOUT     ///< deadline was reached.
+  };
+
+  /// Read from the queue, blocking until an event is available or the queue is
+  /// shutting down.
+  ///
+  /// \param tag[out] Updated to point to the read event's tag.
+  /// \param ok[out] true if read a successful event, false otherwise.
+  ///
+  /// Note that each tag sent to the completion queue (through RPC operations
+  /// or alarms) will be delivered out of the completion queue by a call to
+  /// Next (or a related method), regardless of whether the operation succeeded
+  /// or not. Success here means that this operation completed in the normal
+  /// valid manner.
+  ///
+  /// Server-side RPC request: \a ok indicates that the RPC has indeed
+  /// been started. If it is false, the server has been Shutdown
+  /// before this particular call got matched to an incoming RPC.
+  ///
+  /// Client-side StartCall/RPC invocation: \a ok indicates that the RPC is
+  /// going to go to the wire. If it is false, it not going to the wire. This
+  /// would happen if the channel is either permanently broken or
+  /// transiently broken but with the fail-fast option. (Note that async unary
+  /// RPCs don't post a CQ tag at this point, nor do client-streaming
+  /// or bidi-streaming RPCs that have the initial metadata corked option set.)
+  ///
+  /// Client-side Write, Client-side WritesDone, Server-side Write,
+  /// Server-side Finish, Server-side SendInitialMetadata (which is
+  /// typically included in Write or Finish when not done explicitly):
+  /// \a ok means that the data/metadata/status/etc is going to go to the
+  /// wire. If it is false, it not going to the wire because the call
+  /// is already dead (i.e., canceled, deadline expired, other side
+  /// dropped the channel, etc).
+  ///
+  /// Client-side Read, Server-side Read, Client-side
+  /// RecvInitialMetadata (which is typically included in Read if not
+  /// done explicitly): \a ok indicates whether there is a valid message
+  /// that got read. If not, you know that there are certainly no more
+  /// messages that can ever be read from this stream. For the client-side
+  /// operations, this only happens because the call is dead. For the
+  /// server-sider operation, though, this could happen because the client
+  /// has done a WritesDone already.
+  ///
+  /// Client-side Finish: \a ok should always be true
+  ///
+  /// Server-side AsyncNotifyWhenDone: \a ok should always be true
+  ///
+  /// Alarm: \a ok is true if it expired, false if it was canceled
+  ///
+  /// \return true if got an event, false if the queue is fully drained and
+  ///         shut down.
+  bool Next(void** tag, bool* ok) {
+    return (AsyncNextInternal(tag,
+                              ok,
+                              g_core_codegen_interface->gpr_inf_future(
+                                  GPR_CLOCK_REALTIME)) != SHUTDOWN);
+  }
+
+  /// Read from the queue, blocking up to \a deadline (or the queue's shutdown).
+  /// Both \a tag and \a ok are updated upon success (if an event is available
+  /// within the \a deadline).  A \a tag points to an arbitrary location usually
+  /// employed to uniquely identify an event.
+  ///
+  /// \param tag[out] Upon sucess, updated to point to the event's tag.
+  /// \param ok[out] Upon sucess, true if a successful event, false otherwise
+  ///        See documentation for CompletionQueue::Next for explanation of ok
+  /// \param deadline[in] How long to block in wait for an event.
+  ///
+  /// \return The type of event read.
+  template <typename T>
+  NextStatus AsyncNext(void** tag, bool* ok, const T& deadline) {
+    TimePoint<T> deadline_tp(deadline);
+    return AsyncNextInternal(tag, ok, deadline_tp.raw_time());
+  }
+
+  /// EXPERIMENTAL
+  /// First executes \a F, then reads from the queue, blocking up to
+  /// \a deadline (or the queue's shutdown).
+  /// Both \a tag and \a ok are updated upon success (if an event is available
+  /// within the \a deadline).  A \a tag points to an arbitrary location usually
+  /// employed to uniquely identify an event.
+  ///
+  /// \param F[in] Function to execute before calling AsyncNext on this queue.
+  /// \param tag[out] Upon sucess, updated to point to the event's tag.
+  /// \param ok[out] Upon sucess, true if read a regular event, false otherwise.
+  /// \param deadline[in] How long to block in wait for an event.
+  ///
+  /// \return The type of event read.
+  template <typename T, typename F>
+  NextStatus DoThenAsyncNext(F&& f, void** tag, bool* ok, const T& deadline) {
+    CompletionQueueTLSCache cache = CompletionQueueTLSCache(this);
+    f();
+    if (cache.Flush(tag, ok)) {
+      return GOT_EVENT;
+    } else {
+      return AsyncNext(tag, ok, deadline);
+    }
+  }
+
+  /// Request the shutdown of the queue.
+  ///
+  /// \warning This method must be called at some point if this completion queue
+  /// is accessed with Next or AsyncNext. \a Next will not return false
+  /// until this method has been called and all pending tags have been drained.
+  /// (Likewise for \a AsyncNext returning \a NextStatus::SHUTDOWN .)
+  /// Only once either one of these methods does that (that is, once the queue
+  /// has been \em drained) can an instance of this class be destroyed.
+  /// Also note that applications must ensure that no work is enqueued on this
+  /// completion queue after this method is called.
+  void Shutdown();
+
+  /// Returns a \em raw pointer to the underlying \a grpc_completion_queue
+  /// instance.
+  ///
+  /// \warning Remember that the returned instance is owned. No transfer of
+  /// owership is performed.
+  grpc_completion_queue* cq() { return cq_; }
+
+ protected:
+  /// Private constructor of CompletionQueue only visible to friend classes
+  CompletionQueue(const grpc_completion_queue_attributes& attributes) {
+    cq_ = g_core_codegen_interface->grpc_completion_queue_create(
+        g_core_codegen_interface->grpc_completion_queue_factory_lookup(
+            &attributes),
+        &attributes,
+        NULL);
+    InitialAvalanching();  // reserve this for the future shutdown
+  }
+
+ private:
+  // Friend synchronous wrappers so that they can access Pluck(), which is
+  // a semi-private API geared towards the synchronous implementation.
+  template <class R>
+  friend class ::grpc::ClientReader;
+  template <class W>
+  friend class ::grpc::ClientWriter;
+  template <class W, class R>
+  friend class ::grpc::ClientReaderWriter;
+  template <class R>
+  friend class ::grpc::ServerReader;
+  template <class W>
+  friend class ::grpc::ServerWriter;
+  template <class W, class R>
+  friend class ::grpc::internal::ServerReaderWriterBody;
+  template <class ServiceType, class RequestType, class ResponseType>
+  friend class ::grpc::internal::RpcMethodHandler;
+  template <class ServiceType, class RequestType, class ResponseType>
+  friend class ::grpc::internal::ClientStreamingHandler;
+  template <class ServiceType, class RequestType, class ResponseType>
+  friend class ::grpc::internal::ServerStreamingHandler;
+  template <class Streamer, bool WriteNeeded>
+  friend class ::grpc::internal::TemplatedBidiStreamingHandler;
+  friend class ::grpc::internal::UnknownMethodHandler;
+  friend class ::grpc::Server;
+  friend class ::grpc::ServerContext;
+  friend class ::grpc::ServerInterface;
+  template <class InputMessage, class OutputMessage>
+  friend class ::grpc::internal::BlockingUnaryCallImpl;
+
+  /// EXPERIMENTAL
+  /// Creates a Thread Local cache to store the first event
+  /// On this completion queue queued from this thread.  Once
+  /// initialized, it must be flushed on the same thread.
+  class CompletionQueueTLSCache {
+   public:
+    CompletionQueueTLSCache(CompletionQueue* cq);
+    ~CompletionQueueTLSCache();
+    bool Flush(void** tag, bool* ok);
+
+   private:
+    CompletionQueue* cq_;
+    bool flushed_;
+  };
+
+  NextStatus AsyncNextInternal(void** tag, bool* ok, gpr_timespec deadline);
+
+  /// Wraps \a grpc_completion_queue_pluck.
+  /// \warning Must not be mixed with calls to \a Next.
+  bool Pluck(internal::CompletionQueueTag* tag) {
+    auto deadline =
+        g_core_codegen_interface->gpr_inf_future(GPR_CLOCK_REALTIME);
+    auto ev = g_core_codegen_interface->grpc_completion_queue_pluck(
+        cq_, tag, deadline, nullptr);
+    bool ok = ev.success != 0;
+    void* ignored = tag;
+    GPR_CODEGEN_ASSERT(tag->FinalizeResult(&ignored, &ok));
+    GPR_CODEGEN_ASSERT(ignored == tag);
+    // Ignore mutations by FinalizeResult: Pluck returns the C API status
+    return ev.success != 0;
+  }
+
+  /// Performs a single polling pluck on \a tag.
+  /// \warning Must not be mixed with calls to \a Next.
+  ///
+  /// TODO: sreek - This calls tag->FinalizeResult() even if the cq_ is already
+  /// shutdown. This is most likely a bug and if it is a bug, then change this
+  /// implementation to simple call the other TryPluck function with a zero
+  /// timeout. i.e:
+  ///      TryPluck(tag, gpr_time_0(GPR_CLOCK_REALTIME))
+  void TryPluck(internal::CompletionQueueTag* tag) {
+    auto deadline = g_core_codegen_interface->gpr_time_0(GPR_CLOCK_REALTIME);
+    auto ev = g_core_codegen_interface->grpc_completion_queue_pluck(
+        cq_, tag, deadline, nullptr);
+    if (ev.type == GRPC_QUEUE_TIMEOUT) return;
+    bool ok = ev.success != 0;
+    void* ignored = tag;
+    // the tag must be swallowed if using TryPluck
+    GPR_CODEGEN_ASSERT(!tag->FinalizeResult(&ignored, &ok));
+  }
+
+  /// Performs a single polling pluck on \a tag. Calls tag->FinalizeResult if
+  /// the pluck() was successful and returned the tag.
+  ///
+  /// This exects tag->FinalizeResult (if called) to return 'false' i.e expects
+  /// that the tag is internal not something that is returned to the user.
+  void TryPluck(internal::CompletionQueueTag* tag, gpr_timespec deadline) {
+    auto ev = g_core_codegen_interface->grpc_completion_queue_pluck(
+        cq_, tag, deadline, nullptr);
+    if (ev.type == GRPC_QUEUE_TIMEOUT || ev.type == GRPC_QUEUE_SHUTDOWN) {
+      return;
+    }
+
+    bool ok = ev.success != 0;
+    void* ignored = tag;
+    GPR_CODEGEN_ASSERT(!tag->FinalizeResult(&ignored, &ok));
+  }
+
+  /// Manage state of avalanching operations : completion queue tags that
+  /// trigger other completion queue operations. The underlying core completion
+  /// queue should not really shutdown until all avalanching operations have
+  /// been finalized. Note that we maintain the requirement that an avalanche
+  /// registration must take place before CQ shutdown (which must be maintained
+  /// elsehwere)
+  void InitialAvalanching() {
+    gpr_atm_rel_store(&avalanches_in_flight_, static_cast<gpr_atm>(1));
+  }
+  void RegisterAvalanching() {
+    gpr_atm_no_barrier_fetch_add(&avalanches_in_flight_,
+                                 static_cast<gpr_atm>(1));
+  }
+  void CompleteAvalanching();
+
+  grpc_completion_queue* cq_;  // owned
+
+  gpr_atm avalanches_in_flight_;
+};
+
+/// A specific type of completion queue used by the processing of notifications
+/// by servers. Instantiated by \a ServerBuilder.
+class ServerCompletionQueue : public CompletionQueue {
+ public:
+  bool IsFrequentlyPolled() { return polling_type_ != GRPC_CQ_NON_LISTENING; }
+
+ private:
+  grpc_cq_polling_type polling_type_;
+  friend class ServerBuilder;
+  /// \param is_frequently_polled Informs the GRPC library about whether the
+  /// server completion queue would be actively polled (by calling Next() or
+  /// AsyncNext()). By default all server completion queues are assumed to be
+  /// frequently polled.
+  ServerCompletionQueue(grpc_cq_polling_type polling_type)
+      : CompletionQueue(grpc_completion_queue_attributes{
+            GRPC_CQ_CURRENT_VERSION, GRPC_CQ_NEXT, polling_type}),
+        polling_type_(polling_type) {}
+};
+
+}  // namespace grpc
+
+#endif  // GRPCPP_IMPL_CODEGEN_COMPLETION_QUEUE_H

patches/grpc/fix_too_early_destory.patch

-diff --git a/include/grpcpp/impl/codegen/completion_queue.h b/include/grpcpp/impl/codegen/completion_queue.h
-index 80c7c41982..3f7d8a7714 100644
---- a/include/grpcpp/impl/codegen/completion_queue.h
-+++ b/include/grpcpp/impl/codegen/completion_queue.h
-@@ -32,6 +32,8 @@
- #ifndef GRPCPP_IMPL_CODEGEN_COMPLETION_QUEUE_H
- #define GRPCPP_IMPL_CODEGEN_COMPLETION_QUEUE_H
- 
-+#include <typeinfo>
-+
- #include <grpc/impl/codegen/atm.h>
- #include <grpcpp/impl/codegen/completion_queue_tag.h>
- #include <grpcpp/impl/codegen/core_codegen_interface.h>
-@@ -106,7 +108,9 @@ class CompletionQueue : private GrpcLibraryCodegen {
- 
-   /// Destructor. Destroys the owned wrapped completion queue / instance.
-   ~CompletionQueue() {
--    g_core_codegen_interface->grpc_completion_queue_destroy(cq_);
-+	if (typeid(*g_core_codegen_interface).hash_code() != typeid(CoreCodegenInterface).hash_code()) {
-+      g_core_codegen_interface->grpc_completion_queue_destroy(cq_);
-+	}
-   }
- 
-   /// Tri-state return for AsyncNext: SHUTDOWN, GOT_EVENT, TIMEOUT.
-diff --git a/include/grpcpp/impl/codegen/grpc_library.h b/include/grpcpp/impl/codegen/grpc_library.h
-index 17c904d71a..a092b2204d 100644
---- a/include/grpcpp/impl/codegen/grpc_library.h
-+++ b/include/grpcpp/impl/codegen/grpc_library.h
-@@ -19,6 +19,8 @@
- #ifndef GRPCPP_IMPL_CODEGEN_GRPC_LIBRARY_H
- #define GRPCPP_IMPL_CODEGEN_GRPC_LIBRARY_H
- 
-+#include <typeinfo>
-+
- #include <grpcpp/impl/codegen/core_codegen_interface.h>
- 
- namespace grpc {
-@@ -47,7 +49,8 @@ class GrpcLibraryCodegen {
-     }
-   }
-   virtual ~GrpcLibraryCodegen() {
--    if (grpc_init_called_) {
-+    if (grpc_init_called_ &&
-+		typeid(*g_glip).hash_code() != typeid(GrpcLibraryInterface).hash_code()) {
-       GPR_CODEGEN_ASSERT(g_glip &&
-                          "gRPC library not initialized. See "
-                          "grpc::internal::GrpcLibraryInitializer.");

patches/grpc/grpc_library.h

+// Copyright (c) 2018 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.
+
+#ifndef GRPCPP_IMPL_CODEGEN_GRPC_LIBRARY_H
+#define GRPCPP_IMPL_CODEGEN_GRPC_LIBRARY_H
+
+#include <typeinfo>
+
+#include <grpcpp/impl/codegen/core_codegen_interface.h>
+
+namespace grpc {
+
+class GrpcLibraryInterface {
+ public:
+  virtual ~GrpcLibraryInterface() = default;
+  virtual void init() = 0;
+  virtual void shutdown() = 0;
+};
+
+/// Initialized by \a grpc::GrpcLibraryInitializer from
+/// <grpcpp/impl/grpc_library.h>
+extern GrpcLibraryInterface* g_glip;
+
+/// Classes that require gRPC to be initialized should inherit from this class.
+class GrpcLibraryCodegen {
+ public:
+  GrpcLibraryCodegen(bool call_grpc_init = true) : grpc_init_called_(false) {
+    if (call_grpc_init) {
+      GPR_CODEGEN_ASSERT(g_glip &&
+                         "gRPC library not initialized. See "
+                         "grpc::internal::GrpcLibraryInitializer.");
+      g_glip->init();
+      grpc_init_called_ = true;
+    }
+  }
+  virtual ~GrpcLibraryCodegen() {
+    if (grpc_init_called_ &&
+        typeid(*g_glip).hash_code() !=
+            typeid(GrpcLibraryInterface).hash_code()) {
+      GPR_CODEGEN_ASSERT(g_glip &&
+                         "gRPC library not initialized. See "
+                         "grpc::internal::GrpcLibraryInitializer.");
+      g_glip->shutdown();
+    }
+  }
+
+ private:
+  bool grpc_init_called_;
+};
+
+}  // namespace grpc
+
+#endif  // GRPCPP_IMPL_CODEGEN_GRPC_LIBRARY_H

python/paddle/fluid/__init__.py

@@ -123,7 +123,7 @@ def __bootstrap__():
     read_env_flags = [
         'use_pinned_memory', 'check_nan_inf', 'benchmark', 'warpctc_dir',
         'eager_delete_scope', 'use_mkldnn', 'initial_cpu_memory_in_mb',
-        'init_allocated_mem', 'free_idle_memory'
+        'init_allocated_mem', 'free_idle_memory', 'paddle_num_threads'
     ]
     if core.is_compiled_with_dist():
         read_env_flags.append('rpc_deadline')

python/paddle/fluid/regularizer.py

@@ -142,14 +142,20 @@ class L2DecayRegularizer(WeightDecayRegularizer):
             dtype="float32", shape=param.shape, lod_level=param.lod_level)
 
         if grad.type == core.VarDesc.VarType.SELECTED_ROWS:
+            idx = block.create_var(
+                dtype="int64",
+                shape=param.shape,
+                type=core.VarDesc.VarType.LOD_TENSOR)
             decay = block.create_var(
                 dtype="float32",
                 shape=param.shape,
                 type=core.VarDesc.VarType.SELECTED_ROWS)
+            block.append_op(
+                type='extract_rows', inputs={'X': grad}, outputs={'Out': idx})
             block.append_op(
                 type='lookup_table',
                 inputs={'W': param,
-                        'Ids': grad},
+                        'Ids': idx},
                 outputs={'Out': decay},
                 attrs={'is_sparse': True})
             param = decay
@@ -216,14 +222,20 @@ class L1DecayRegularizer(WeightDecayRegularizer):
             dtype="float32", shape=param.shape, lod_level=param.lod_level)
 
         if grad.type == core.VarDesc.VarType.SELECTED_ROWS:
+            idx = block.create_var(
+                dtype="int64",
+                shape=param.shape,
+                type=core.VarDesc.VarType.LOD_TENSOR)
             decay = block.create_var(
                 dtype="float32",
                 shape=param.shape,
                 type=core.VarDesc.VarType.SELECTED_ROWS)
+            block.append_op(
+                type='extract_rows', inputs={'X': grad}, outputs={'Out': idx})
             block.append_op(
                 type='lookup_table',
                 inputs={'W': param,
-                        'Ids': grad},
+                        'Ids': idx},
                 outputs={'Out': decay},
                 attrs={'is_sparse': True})
 

python/paddle/fluid/tests/unittests/dist_se_resnext.py

@@ -278,7 +278,7 @@ class DistSeResneXt2x2:
 
     def run_trainer(self, place, endpoints, trainer_id, trainers, is_dist=True):
         test_program, avg_cost, train_reader, test_reader, batch_acc, predict = get_model(
-            batch_size=20)
+            batch_size=2)
         if is_dist:
             t = get_transpiler(trainer_id,
                                fluid.default_main_program(), endpoints,
@@ -294,11 +294,7 @@ class DistSeResneXt2x2:
         strategy.num_threads = 1
         strategy.allow_op_delay = False
         exe = fluid.ParallelExecutor(
-            True,
-            loss_name=avg_cost.name,
-            exec_strategy=strategy,
-            num_trainers=trainers,
-            trainer_id=trainer_id)
+            True, loss_name=avg_cost.name, exec_strategy=strategy)
 
         feed_var_list = [
             var for var in trainer_prog.global_block().vars.itervalues()

python/paddle/fluid/tests/unittests/test_dist_se_resnext.py

@@ -56,7 +56,7 @@ class TestDistSeResneXt2x2(unittest.TestCase):
             except os.error:
                 retry_times -= 1
 
-    def non_test_with_place(self):
+    def no_test_with_place(self):
         # *ATTENTION* THIS TEST NEEDS AT LEAST 2GPUS TO RUN
         required_envs = {
             "PATH": os.getenv("PATH"),

python/paddle/fluid/tests/unittests/test_extract_rows_op.py

+#   Copyright (c) 2018 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.
+
+import unittest
+import numpy as np
+import paddle.fluid.core as core
+from paddle.fluid.op import Operator
+from op_test import OpTest
+
+
+class TestExtractRows(OpTest):
+    def check_with_place(self, place):
+        scope = core.Scope()
+
+        # create and initialize Variable
+        feature_len = 12
+        rows = [0, 4, 4, 7]
+        np_array = np.ones((len(rows), feature_len)).astype("float32")
+
+        in_x = scope.var('X').get_selected_rows()
+        in_x.set_height(len(rows))
+        in_x.set_rows(rows)
+        in_x_tensor = in_x.get_tensor()
+        in_x_tensor.set(np_array, place)
+
+        # create Out Variable
+        out_tensor = scope.var('Out').get_tensor()
+
+        # create and run lookup_table operator
+        extract_rows_op = Operator("extract_rows", X='X', Out='Out')
+        extract_rows_op.run(scope, place)
+
+        # get result from Out
+        result_array = np.array(out_tensor)
+        result_array = [ele[0] for ele in result_array]
+        assert result_array == rows
+
+    def test_concat_rows(self):
+        places = [core.CPUPlace()]
+        if core.is_compiled_with_cuda():
+            places.append(core.CUDAPlace(0))
+        for place in places:
+            self.check_with_place(place)
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/test_flatten_op.py

+#   Copyright (c) 2018 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.
+
+import unittest
+import numpy as np
+
+from op_test import OpTest
+
+
+class TestFlattenOp(OpTest):
+    def setUp(self):
+        self.op_type = "flatten"
+        self.init_test_case()
+        self.inputs = {"X": np.random.random(self.in_shape).astype("float32")}
+        self.init_attrs()
+        self.outputs = {"Out": self.inputs["X"].reshape(self.new_shape)}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(["X"], "Out")
+
+    def init_test_case(self):
+        self.in_shape = (3, 2, 2, 5)
+        self.axis = 1
+        self.new_shape = (3, 20)
+
+    def init_attrs(self):
+        self.attrs = {"axis": self.axis}
+
+
+class TestFlattenOp(TestFlattenOp):
+    def init_test_case(self):
+        self.in_shape = (3, 2, 2, 3)
+        self.axis = 0
+        self.new_shape = (1, 36)
+
+
+class TestFlattenOpWithDefaultAxis(TestFlattenOp):
+    def init_test_case(self):
+        self.in_shape = (3, 2, 2, 3)
+        self.new_shape = (3, 12)
+
+    def init_attrs(self):
+        self.attrs = {}
+
+
+class TestFlattenOpSixDims(TestFlattenOp):
+    def init_test_case(self):
+        self.in_shape = (3, 2, 3, 2, 4, 4)
+        self.axis = 4
+        self.new_shape = (36, 16)
+
+
+if __name__ == "__main__":
+    unittest.main()

python/paddle/fluid/tests/unittests/test_lookup_table_op.py

@@ -49,53 +49,6 @@ class TestLookupTableOpWithPadding(TestLookupTableOp):
         pass
 
 
-class TestLookupTableIdsIsSelectedRows(OpTest):
-    def check_with_place(self, place):
-        scope = core.Scope()
-
-        # create and initialize Variable
-        height = 10
-        rows = [0, 4, 4, 7]
-        row_numel = 12
-
-        # create and initialize W Variable
-        W = scope.var('W').get_tensor()
-        W_array = np.full((height, row_numel), 1.0).astype("float32")
-        for i in range(height):
-            W_array[i] *= i
-        W.set(W_array, place)
-
-        # create and initialize Ids Variable
-        ids_selected_rows = scope.var('Ids').get_selected_rows()
-        ids_selected_rows.set_height(len(rows))
-        ids_selected_rows.set_rows(rows)
-        np_array = np.ones((len(rows), row_numel)).astype("float32")
-        ids_tensor = ids_selected_rows.get_tensor()
-        ids_tensor.set(np_array, place)
-
-        # create Out Variable
-        Out = scope.var('Out').get_selected_rows()
-
-        # create and run lookup_table operator
-        concat_rows_op = Operator("lookup_table", W='W', Ids='Ids', Out='Out')
-        concat_rows_op.run(scope, place)
-
-        # get result from Out
-        Out_tensor = Out.get_tensor()
-        result_array = np.array(Out_tensor)
-
-        # all(): return True if all elements of the iterable are true (or if the iterable is empty)
-        for idx, row in enumerate(rows):
-            assert (row == result_array[idx]).all()
-
-    def test_concat_rows(self):
-        places = [core.CPUPlace()]
-        if core.is_compiled_with_cuda():
-            places.append(core.CUDAPlace(0))
-        for place in places:
-            self.check_with_place(place)
-
-
 class TestLookupTableWIsSelectedRows(OpTest):
     def check_with_place(self, place):
         scope = core.Scope()

tools/codestyle/cpplint_pre_commit.hook

@@ -4,7 +4,7 @@ TOTAL_ERRORS=0
 
 # The trick to remove deleted files: https://stackoverflow.com/a/2413151
 for file in $(git diff --cached --name-status | awk '$1 != "D" {print $2}'); do
-    if [[ $file =~ ^(paddle/legacy/api/.*|paddle/legacy/capi/.*|paddle/contrib/.*|paddle/legacy/cuda/.*|paddle/legacy/function/.*|paddle/legacy/gserver/.*|paddle/legacy/math/.*|paddle/legacy/optimizer/.*|paddle/legacy/parameter/.*|paddle/legacy/pserver/.*|paddle/legacy/trainer/.*|paddle/legacy/utils/.*|paddle/testing/TestUtil.*) ]]; then
+    if [[ $file =~ ^(paddle/legacy/api/.*|paddle/legacy/capi/.*|paddle/contrib/.*|paddle/legacy/cuda/.*|paddle/legacy/function/.*|paddle/legacy/gserver/.*|paddle/legacy/math/.*|paddle/legacy/optimizer/.*|paddle/legacy/parameter/.*|paddle/legacy/pserver/.*|paddle/legacy/trainer/.*|paddle/legacy/utils/.*|paddle/testing/TestUtil.*|patches/grpc/.*) ]]; then
         continue;
     else
         cpplint --filter=-readability/fn_size $file;