merge develop

test=develop

CMakeLists.txt

@@ -54,7 +54,7 @@ option(WITH_PYTHON      "Compile PaddlePaddle with python interpreter"  ON)
 option(WITH_DOUBLE      "Compile PaddlePaddle with double precision"    OFF)
 option(WITH_RDMA        "Compile PaddlePaddle with RDMA support"        OFF)
 option(WITH_TIMER       "Compile PaddlePaddle with stats timer"         OFF)
-option(WITH_PROFILER    "Compile PaddlePaddle with GPU profiler"        OFF)
+option(WITH_PROFILER    "Compile PaddlePaddle with GPU profiler and gperftools"        OFF)
 option(WITH_DOC         "Compile PaddlePaddle with documentation"       OFF)
 option(WITH_COVERAGE    "Compile PaddlePaddle with code coverage"       OFF)
 option(COVERALLS_UPLOAD "Package code coverage data to coveralls"       OFF)
@@ -254,6 +254,12 @@ elseif()
     set(WITH_ANAKIN OFF CACHE STRING "Anakin is used in MKL only now." FORCE)
 endif()
 
+if (WITH_PROFILER)
+    find_package(Gperftools REQUIRED)
+    include_directories(${GPERFTOOLS_INCLUDE_DIR})
+    add_definitions(-DWITH_GPERFTOOLS)
+endif()
+
 include(generic)            # simplify cmake module
 include(package)            # set paddle packages
 include(ccache)             # set ccache for compilation

README.md

@@ -2,8 +2,8 @@
 
 
 [![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
-[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.1/getstarted/index_en.html)
-[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html)
+[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)
+[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html)
 [![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
 [![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
 
@@ -19,7 +19,7 @@ Our vision is to enable deep learning for everyone via PaddlePaddle.
 Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddle/releases) to track the latest feature of PaddlePaddle.
 
 
-### Latest PaddlePaddle Release: [Fluid 1.1.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.1)
+### Latest PaddlePaddle Release: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2)
 ### Install Latest Stable Release:
 ```
 # Linux CPU
@@ -27,9 +27,9 @@ pip install paddlepaddle
 # Linux GPU cuda9cudnn7
 pip install paddlepaddle-gpu
 # Linux GPU cuda8cudnn7
-pip install paddlepaddle-gpu==1.1.0.post87
+pip install paddlepaddle-gpu==1.2.0.post87
 # Linux GPU cuda8cudnn5
-pip install paddlepaddle-gpu==1.1.0.post85
+pip install paddlepaddle-gpu==1.2.0.post85
 
 # For installation on other platform, refer to http://paddlepaddle.org/
 ```
@@ -76,26 +76,26 @@ pip install paddlepaddle-gpu==1.1.0.post85
 
 ## Installation
 
-It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html) on our website.
+It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html) on our website.
 
 ## Documentation
 
-We provide [English](http://paddlepaddle.org/documentation/docs/en/1.1/getstarted/index_en.html) and
-[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html) documentation.
+We provide [English](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html) and
+[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html) documentation.
 
 - [Deep Learning 101](https://github.com/PaddlePaddle/book)
 
   You might want to start from this online interactive book that can run in a Jupyter Notebook.
 
-- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/1.1/user_guides/howto/training/cluster_howto.html)
+- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/1.2/user_guides/howto/training/cluster_howto.html)
 
   You can run distributed training jobs on MPI clusters.
 
-- [Python API](http://paddlepaddle.org/documentation/api/zh/1.1/fluid.html)
+- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.2/api_cn/index_cn.html)
 
    Our new API enables much shorter programs.
 
-- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.1/advanced_usage/development/contribute_to_paddle.html)
+- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.2/advanced_usage/development/contribute_to_paddle/index_cn.html)
 
    We appreciate your contributions!
 

cmake/FindGperftools.cmake

+# Tries to find Gperftools.
+#
+# Usage of this module as follows:
+#
+#     find_package(Gperftools)
+#
+# Variables used by this module, they can change the default behaviour and need
+# to be set before calling find_package:
+#
+#  Gperftools_ROOT_DIR  Set this variable to the root installation of
+#                       Gperftools if the module has problems finding
+#                       the proper installation path.
+#
+# Variables defined by this module:
+#
+#  GPERFTOOLS_FOUND              System has Gperftools libs/headers
+#  GPERFTOOLS_LIBRARIES          The Gperftools libraries (tcmalloc & profiler)
+#  GPERFTOOLS_INCLUDE_DIR        The location of Gperftools headers
+
+find_library(GPERFTOOLS_TCMALLOC
+  NAMES tcmalloc
+  HINTS ${Gperftools_ROOT_DIR}/lib)
+
+find_library(GPERFTOOLS_PROFILER
+  NAMES profiler
+  HINTS ${Gperftools_ROOT_DIR}/lib)
+
+find_library(GPERFTOOLS_TCMALLOC_AND_PROFILER
+  NAMES tcmalloc_and_profiler
+  HINTS ${Gperftools_ROOT_DIR}/lib)
+
+find_path(GPERFTOOLS_INCLUDE_DIR
+  NAMES gperftools/heap-profiler.h
+  HINTS ${Gperftools_ROOT_DIR}/include)
+
+set(GPERFTOOLS_LIBRARIES ${GPERFTOOLS_TCMALLOC_AND_PROFILER})
+
+include(FindPackageHandleStandardArgs)
+find_package_handle_standard_args(
+  Gperftools
+  DEFAULT_MSG
+  GPERFTOOLS_LIBRARIES
+  GPERFTOOLS_INCLUDE_DIR)
+
+mark_as_advanced(
+  Gperftools_ROOT_DIR
+  GPERFTOOLS_TCMALLOC
+  GPERFTOOLS_PROFILER
+  GPERFTOOLS_TCMALLOC_AND_PROFILER
+  GPERFTOOLS_LIBRARIES
+  GPERFTOOLS_INCLUDE_DIR)
+
+# create IMPORTED targets
+if (Gperftools_FOUND AND NOT TARGET gperftools::tcmalloc)
+  add_library(gperftools::tcmalloc UNKNOWN IMPORTED)
+  set_target_properties(gperftools::tcmalloc PROPERTIES
+    IMPORTED_LOCATION ${GPERFTOOLS_TCMALLOC}
+    INTERFACE_INCLUDE_DIRECTORIES "${GPERFTOOLS_INCLUDE_DIR}")
+  add_library(gperftools::profiler UNKNOWN IMPORTED)
+  set_target_properties(gperftools::profiler PROPERTIES
+    IMPORTED_LOCATION ${GPERFTOOLS_PROFILER}
+    INTERFACE_INCLUDE_DIRECTORIES "${GPERFTOOLS_INCLUDE_DIR}")
+endif()

cmake/configure.cmake

@@ -86,6 +86,7 @@ endif(NOT WITH_GOLANG)
 
 if(WITH_GPU)
     add_definitions(-DPADDLE_WITH_CUDA)
+    add_definitions(-DEIGEN_USE_GPU)
 
     FIND_PACKAGE(CUDA REQUIRED)
 

cmake/generic.cmake

@@ -110,6 +110,14 @@ function(find_fluid_modules TARGET_NAME)
   endif()
 endfunction(find_fluid_modules)
 
+
+function(common_link TARGET_NAME)
+  if (WITH_PROFILER)
+    target_link_libraries(${TARGET_NAME} gperftools::profiler)
+  endif()
+endfunction()
+
+
 # find all third_party modules is used for paddle static library
 # for reduce the dependency when building the inference libs.
 set_property(GLOBAL PROPERTY FLUID_THIRD_PARTY)
@@ -274,6 +282,7 @@ function(cc_library TARGET_NAME)
       endif()
       target_link_libraries(${TARGET_NAME} ${cc_library_DEPS})
       add_dependencies(${TARGET_NAME} ${cc_library_DEPS})
+      common_link(${TARGET_NAME})
     endif()
 
     # cpplint code style
@@ -340,6 +349,7 @@ function(cc_binary TARGET_NAME)
   if(cc_binary_DEPS)
     target_link_libraries(${TARGET_NAME} ${cc_binary_DEPS})
     add_dependencies(${TARGET_NAME} ${cc_binary_DEPS})
+    common_link(${TARGET_NAME})
   endif()
 endfunction(cc_binary)
 
@@ -362,6 +372,7 @@ function(cc_test TARGET_NAME)
       target_link_libraries(${TARGET_NAME} ${win32_deps})
     endif(WIN32)
     add_dependencies(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
+    common_link(${TARGET_NAME})
     add_test(NAME ${TARGET_NAME}
              COMMAND ${TARGET_NAME} ${cc_test_ARGS}
              WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
@@ -420,6 +431,7 @@ function(nv_binary TARGET_NAME)
     if(nv_binary_DEPS)
       target_link_libraries(${TARGET_NAME} ${nv_binary_DEPS})
       add_dependencies(${TARGET_NAME} ${nv_binary_DEPS})
+      common_link(${TARGET_NAME})
     endif()
   endif()
 endfunction(nv_binary)
@@ -433,6 +445,7 @@ function(nv_test TARGET_NAME)
     cuda_add_executable(${TARGET_NAME} ${nv_test_SRCS})
     target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
     add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
+    common_link(${TARGET_NAME})
     add_test(${TARGET_NAME} ${TARGET_NAME})
     if (nv_test_SERIAL)
         set_property(TEST ${TARGET_NAME} PROPERTY RUN_SERIAL 1)
@@ -499,6 +512,7 @@ function(hip_binary TARGET_NAME)
     if(hip_binary_DEPS)
       target_link_libraries(${TARGET_NAME} ${hip_binary_DEPS})
       add_dependencies(${TARGET_NAME} ${hip_binary_DEPS})
+      common_link(${TARGET_NAME})
     endif()
   endif()
 endfunction(hip_binary)
@@ -518,6 +532,7 @@ function(hip_test TARGET_NAME)
     set_target_properties(${TARGET_NAME} PROPERTIES LINKER_LANGUAGE HIP)
     target_link_libraries(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main memory gtest gflags)
     add_dependencies(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main memory gtest gflags)
+    common_link(${TARGET_NAME})
     add_test(${TARGET_NAME} ${TARGET_NAME})
   endif()
 endfunction(hip_test)
@@ -560,6 +575,7 @@ function(go_library TARGET_NAME)
   endif()
   if(go_library_DEPS)
     add_dependencies(${TARGET_NAME} ${go_library_DEPS})
+    common_link(${TARGET_NAME})
   endif(go_library_DEPS)
 
   # The "source file" of the library is `${dummyfile}` which never

paddle/fluid/API.spec

@@ -66,6 +66,7 @@ paddle.fluid.layers.linear_chain_crf ArgSpec(args=['input', 'label', 'param_attr
 paddle.fluid.layers.crf_decoding ArgSpec(args=['input', 'param_attr', 'label'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.cos_sim ArgSpec(args=['X', 'Y'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.cross_entropy ArgSpec(args=['input', 'label', 'soft_label', 'ignore_index'], varargs=None, keywords=None, defaults=(False, -100))
+paddle.fluid.layers.bpr_loss ArgSpec(args=['input', 'label', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.square_error_cost ArgSpec(args=['input', 'label'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.chunk_eval ArgSpec(args=['input', 'label', 'chunk_scheme', 'num_chunk_types', 'excluded_chunk_types'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.sequence_conv ArgSpec(args=['input', 'num_filters', 'filter_size', 'filter_stride', 'padding', 'bias_attr', 'param_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(3, 1, None, None, None, None, None))

paddle/fluid/CMakeLists.txt

 add_subdirectory(memory)
 add_subdirectory(platform)
 add_subdirectory(framework)
+add_subdirectory(imperative)
 add_subdirectory(operators)
 add_subdirectory(string)
 add_subdirectory(recordio)

paddle/fluid/framework/CMakeLists.txt

@@ -129,11 +129,13 @@ cc_test(version_test SRCS version_test.cc DEPS version)
 
 cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog version)
 
-if(NOT WIN32)
-cc_library(ngraph_bridge SRCS ngraph_bridge.cc DEPS operator framework_proto ngraph)
-cc_library(ngraph_operator SRCS ngraph_operator.cc DEPS ngraph_bridge operator op_info device_context tensor scope glog
-  shape_inference data_transform lod_tensor profiler)
-endif(NOT WIN32)
+if(WITH_NGRAPH)
+  if(NOT WIN32)
+    cc_library(ngraph_bridge SRCS ngraph_bridge.cc DEPS operator framework_proto ngraph)
+    cc_library(ngraph_operator SRCS ngraph_operator.cc DEPS ngraph_bridge operator op_info device_context tensor scope glog
+      shape_inference data_transform lod_tensor profiler ngraph)
+  endif(NOT WIN32)
+endif(WITH_NGRAPH)
 
 cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc)
 nv_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
@@ -169,11 +171,15 @@ if(WITH_DISTRIBUTE)
   set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
   set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
 else()
-  if(NOT WIN32)
-    cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass ngraph_operator variable_helper)
-  else(NOT WIN32)
+  if(WITH_NGRAPH)
+    if(NOT WIN32)
+      cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass ngraph ngraph_operator variable_helper)
+    else(NOT WIN32)
+      cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper)
+    endif(NOT WIN32)
+  else(WITH_NGRAPH)
     cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper)
-  endif(NOT WIN32)
+  endif(WITH_NGRAPH)
   cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op)
 endif()
 

paddle/fluid/framework/data_layout_transform.cc

@@ -151,19 +151,22 @@ void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var,
   auto out_format =
       platform::MKLDNNFormatForSize(in_tz.size(), ToMKLDNNFormat(out_layout));
 
-  void* in_data = GetDataFromTensor(in, in_type);
-
   // output tensor has the same dims as input. Reorder don't change dims
   out->Resize(in.dims());
 
-  auto out_data = out->mutable_data(expected_kernel_type.place_, in.type());
-
-  auto in_memory = memory({{{in_tz}, in_type, in_format}, cpu_engine}, in_data);
-  auto out_memory =
-      memory({{{out_tz}, out_type, out_format}, cpu_engine}, out_data);
+  if (in_format != out_format) {
+    void* in_data = GetDataFromTensor(in, in_type);
+    auto out_data = out->mutable_data(expected_kernel_type.place_, in.type());
 
-  platform::Reorder(in_memory, out_memory);
+    auto in_memory =
+        memory({{{in_tz}, in_type, in_format}, cpu_engine}, in_data);
+    auto out_memory =
+        memory({{{out_tz}, out_type, out_format}, cpu_engine}, out_data);
 
+    platform::Reorder(in_memory, out_memory);
+  } else {
+    out->ShareDataWith(in);
+  }
   out->set_layout(out_layout);
   // reset format since the out tensor will be feed to non-MKLDNN OPkernel
   out->set_format(memory::format::format_undef);

paddle/fluid/framework/executor.cc

@@ -17,7 +17,6 @@ limitations under the License. */
 #include "paddle/fluid/framework/feed_fetch_method.h"
 #include "paddle/fluid/framework/lod_rank_table.h"
 #include "paddle/fluid/framework/lod_tensor_array.h"
-#include "paddle/fluid/framework/ngraph_operator.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/reader.h"
 #include "paddle/fluid/framework/transfer_scope_cache.h"
@@ -26,6 +25,10 @@ limitations under the License. */
 #include "paddle/fluid/platform/place.h"
 #include "paddle/fluid/platform/profiler.h"
 
+#ifdef PADDLE_WITH_NGRAPH
+#include "paddle/fluid/framework/ngraph_operator.h"
+#endif
+
 DECLARE_bool(benchmark);
 DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run");
 DEFINE_bool(use_ngraph, false, "Use NGRAPH to run");
@@ -88,11 +91,11 @@ static void DeleteUnusedTensors(const Scope& scope, const OperatorBase* op,
 static void EnableFusedOp(ExecutorPrepareContext* ctx) {
 #ifdef PADDLE_WITH_NGRAPH
   VLOG(3) << "use_ngraph=True";
-  auto intervals = FusedOperator::FusedOpIntervals(&ctx->ops_);
+  auto intervals = NgraphOperator::NgraphOpIntervals(&ctx->ops_);
   for (auto& interval : intervals) {
-    auto* fused_op = new FusedOperator(ctx->prog_, ctx->block_id_,
-                                       interval.at(0), interval.at(1));
-    *interval[0] = std::unique_ptr<OperatorBase>(fused_op);
+    auto* ng_op = new NgraphOperator(ctx->prog_, ctx->block_id_, interval.at(0),
+                                     interval.at(1));
+    *interval[0] = std::unique_ptr<OperatorBase>(ng_op);
   }
   for (auto it = intervals.rbegin(); it != intervals.rend(); ++it) {
     ctx->ops_.erase(it->at(0) + 1, it->at(1));

paddle/fluid/framework/feed_fetch_method.cc

@@ -16,7 +16,9 @@ limitations under the License. */
 #include <string>
 #include <vector>
 #include "glog/logging.h"
+#include "paddle/fluid/framework/var_type.h"
 #include "paddle/fluid/framework/variable.h"
+#include "paddle/fluid/platform/place.h"
 
 namespace paddle {
 namespace framework {
@@ -53,5 +55,12 @@ LoDTensor& GetFetchVariable(const Scope& scope, const std::string& var_name,
   return tensor;
 }
 
+LoDTensor& GetVariableTensor(const Scope& scope, const std::string& var_name) {
+  Variable* var = scope.FindVar(var_name);
+  PADDLE_ENFORCE(var, "%s no in scope", var_name);
+  PADDLE_ENFORCE(var->IsType<LoDTensor>(), "Only support lod tensor now.");
+  return *var->GetMutable<LoDTensor>();
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/feed_fetch_method.h

@@ -27,5 +27,7 @@ void SetFeedVariable(Scope* scope, const LoDTensor& input,
 LoDTensor& GetFetchVariable(const Scope& scope, const std::string& var_name,
                             size_t index);
 
+LoDTensor& GetVariableTensor(const Scope& scope, const std::string& var_name);
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/ir/graph.cc

@@ -38,9 +38,8 @@ void CheckProgram(const ProgramDesc &program) {
     switch (role_id) {
       case _INT(OpRole::kForward):
         if (visit.find(_INT(OpRole::kBackward)) != visit.end()) {
-          LOG(ERROR)
-              << "Cannot add backward operator before forward operator %s."
-              << op->Type();
+          LOG(ERROR) << "Cannot add backward operator before forward operator "
+                     << op->Type();
         }
         break;
       case _INT(OpRole::kBackward):

paddle/fluid/framework/ir/is_test_pass.cc

@@ -38,7 +38,7 @@ std::unique_ptr<ir::Graph> IsTestPass::ApplyImpl(
   for (const Node* n : graph->Nodes()) {
     if (n->IsOp()) {
       auto* op = n->Op();
-      if (n->RuntimeHasAttr("is_test")) {
+      if (op->HasAttr("is_test") || op->HasProtoAttr("is_test")) {
         op->SetAttr("is_test", true);
       } else if (std::find(begin(op_list), end(op_list), op->Type()) !=
                  end(op_list)) {

paddle/fluid/framework/ir/is_test_pass_tester.cc

@@ -104,9 +104,9 @@ TEST(IsTestPass, basic) {
       auto* op = node->Op();
       auto op_name = boost::get<std::string>(op->GetAttr("name"));
       if (op_name == "conv3") {
-        ASSERT_FALSE(node->RuntimeHasAttr("is_test"));
+        ASSERT_FALSE(op->HasAttr("is_test"));
       } else {
-        ASSERT_TRUE(node->RuntimeHasAttr("is_test"));
+        ASSERT_TRUE(op->HasAttr("is_test"));
         EXPECT_TRUE(boost::get<bool>(op->GetAttr("is_test")));
       }
     }

paddle/fluid/framework/ir/mkldnn_placement_pass.cc

@@ -25,12 +25,15 @@ std::unique_ptr<ir::Graph> MKLDNNPlacementPass::ApplyImpl(
   const auto& op_types_list =
       Get<std::unordered_set<std::string>>("mkldnn_enabled_op_types");
   for (const Node* n : graph->Nodes()) {
-    if (n->IsOp() && n->RuntimeHasAttr("use_mkldnn")) {
-      if (op_types_list.empty()) {
-        n->Op()->SetAttr("use_mkldnn", true);
-      } else if (std::find(op_types_list.begin(), op_types_list.end(),
-                           n->Name()) != op_types_list.end()) {
-        n->Op()->SetAttr("use_mkldnn", true);
+    if (n->IsOp()) {
+      auto* op = n->Op();
+      if (op->HasAttr("use_mkldnn") || op->HasProtoAttr("use_mkldnn")) {
+        if (op_types_list.empty()) {
+          op->SetAttr("use_mkldnn", true);
+        } else if (std::find(op_types_list.begin(), op_types_list.end(),
+                             n->Name()) != op_types_list.end()) {
+          op->SetAttr("use_mkldnn", true);
+        }
       }
     }
   }

paddle/fluid/framework/ir/node.cc

@@ -30,28 +30,6 @@ std::unique_ptr<Node> CreateNodeForTest(const std::string &name,
   return std::unique_ptr<Node>(new Node(name, type));
 }
 
-bool Node::RuntimeHasAttr(const std::string &name) const {
-  if (Op()->HasAttr(name)) {
-    return true;
-  } else {
-    auto &op_info = OpInfoMap::Instance();
-    auto op_type = Op()->Type();
-    if (op_info.Has(op_type)) {
-      auto op_info_ptr = op_info.Get(op_type);
-      if (op_info_ptr.HasOpProtoAndChecker()) {
-        const proto::OpProto &proto = op_info_ptr.Proto();
-        for (int i = 0; i != proto.attrs_size(); ++i) {
-          const proto::OpProto::Attr &attr = proto.attrs(i);
-          if (attr.name() == name) {
-            return true;
-          }
-        }
-      }
-    }
-  }
-  return false;
-}
-
 }  // namespace ir
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/ir/node.h

@@ -108,18 +108,6 @@ class Node {
            Name().find(ir::Node::kControlDepVarName) != std::string::npos;
   }
 
-  // RuntimeHasAttr is different with HasAttr now.
-  // 1. For Op()->HasAttr(), it judges whether a stored program_desc_ has attr,
-  // thus, if stored program_desc_ are old which don't have an attr, a new
-  // library which adds the attr already will fail on this function.
-  // Details:
-  // https://github.com/PaddlePaddle/Paddle/pull/14608#issuecomment-442309087
-  // 2. For Op()->RuntimeHasAttr, it judges the attr in runtime to avoid above
-  // problem.
-  // TODO(luotao): Maybe we should enhance HasAttr later, instead of adding
-  // RuntimeHasAttr.
-  bool RuntimeHasAttr(const std::string& name) const;
-
   std::vector<Node*> inputs;
   std::vector<Node*> outputs;
 

paddle/fluid/framework/ngraph_bridge.cc

@@ -12,7 +12,6 @@ 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. */
 
-#ifdef PADDLE_WITH_NGRAPH
 #include <algorithm>
 #include <functional>
 #include <vector>
@@ -27,14 +26,15 @@ namespace paddle {
 namespace framework {
 
 static std::shared_ptr<ngraph::Node> GetNode(
-    const std::shared_ptr<OperatorBase>& op, const std::string prm,
+    const std::shared_ptr<OperatorBase>& op, const std::string name,
     const VariableNameMap& var_map,
     std::shared_ptr<
         std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
         ngb_node_map) {
-  auto& var_names = var_map.at(prm);
+  auto& var_names = var_map.at(name);
   PADDLE_ENFORCE_EQ(var_names.size(), 1,
-                    "op %s prm %s expects one associated var", op->Type(), prm);
+                    "op %s name %s expects one associated var", op->Type(),
+                    name);
   if (ngb_node_map->find(var_names[0]) != ngb_node_map->end()) {
     return (*ngb_node_map)[var_names[0]];
   } else {
@@ -43,42 +43,42 @@ static std::shared_ptr<ngraph::Node> GetNode(
 }
 
 static std::shared_ptr<ngraph::Node> GetInputNode(
-    const std::shared_ptr<OperatorBase>& op, const std::string prm,
+    const std::shared_ptr<OperatorBase>& op, const std::string name,
     std::shared_ptr<
         std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
         ngb_node_map) {
-  return GetNode(op, prm, op->Inputs(), ngb_node_map);
+  return GetNode(op, name, op->Inputs(), ngb_node_map);
 }
 
 static std::shared_ptr<ngraph::Node> GetOutputNode(
-    const std::shared_ptr<OperatorBase>& op, const std::string prm,
+    const std::shared_ptr<OperatorBase>& op, const std::string name,
     std::shared_ptr<
         std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
         ngb_node_map) {
-  return GetNode(op, prm, op->Outputs(), ngb_node_map);
+  return GetNode(op, name, op->Outputs(), ngb_node_map);
 }
 
 static void SetOutputNode(
-    const std::shared_ptr<OperatorBase>& op, const std::string prm,
+    const std::shared_ptr<OperatorBase>& op, const std::string name,
     std::shared_ptr<ngraph::Node> node,
     std::shared_ptr<
         std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
         ngb_node_map) {
-  auto& var_names = op->Outputs().at(prm);
+  auto& var_names = op->Outputs().at(name);
   if (var_names.size() == 1) {
     (*ngb_node_map)[var_names[0]] = node;
   } else if (var_names.size() == 0) {
     (*ngb_node_map)[""] = node;
   } else {
-    PADDLE_THROW("prm %s has more than 1 var_names.", prm);
+    PADDLE_THROW("name %s has more than 1 var_names.", name);
   }
 }
 
 static bool HasOutput(const std::shared_ptr<OperatorBase>& op,
-                      const std::string prm) {
+                      const std::string name) {
   auto& outputs = op->Outputs();
-  if (outputs.find(prm) == outputs.end()) return false;
-  return outputs.at(prm).size() > 0;
+  if (outputs.find(name) == outputs.end()) return false;
+  return outputs.at(name).size() > 0;
 }
 
 template <typename T>
@@ -118,4 +118,3 @@ void NgraphBridge::BuildNgNode(const std::shared_ptr<OperatorBase>& op) {
 
 }  // namespace framework
 }  // namespace paddle
-#endif

paddle/fluid/framework/ngraph_bridge.h

@@ -14,8 +14,6 @@ limitations under the License. */
 
 #pragma once
 
-#ifdef PADDLE_WITH_NGRAPH
-
 #include <algorithm>
 #include <map>
 #include <string>
@@ -53,4 +51,3 @@ class NgraphBridge {
 
 }  // namespace framework
 }  // namespace paddle
-#endif

paddle/fluid/framework/ngraph_operator.cc

@@ -12,7 +12,6 @@ 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. */
 
-#ifdef PADDLE_WITH_NGRAPH
 #include <glog/logging.h>
 
 #include <algorithm>
@@ -58,16 +57,16 @@ typedef enum {                /* nGraph support state on ops          */
 } op_state;
 
 // perform graph build through bridge and execute computation
-class NgraphOperator {
+class NgraphEngine {
  public:
-  explicit NgraphOperator(const Scope& scope, const platform::Place& place,
-                          const std::vector<std::shared_ptr<OperatorBase>>& ops,
-                          const std::unordered_map<
-                              std::string, ngraph::element::Type>& var_type_map,
-                          const std::unordered_set<std::string>& persist,
-                          const std::unordered_set<std::string>& fetches,
-                          const std::unordered_set<std::string>& post_op_inputs,
-                          op_state ng_op_state)
+  explicit NgraphEngine(const Scope& scope, const platform::Place& place,
+                        const std::vector<std::shared_ptr<OperatorBase>>& ops,
+                        const std::unordered_map<
+                            std::string, ngraph::element::Type>& var_type_map,
+                        const std::unordered_set<std::string>& persist,
+                        const std::unordered_set<std::string>& fetches,
+                        const std::unordered_set<std::string>& post_op_inputs,
+                        op_state ng_op_state)
       : scope_(scope),
         place_(place),
         fused_ops_(ops),
@@ -132,7 +131,7 @@ class NgraphOperator {
 };
 
 std::vector<std::vector<std::vector<std::unique_ptr<OperatorBase>>::iterator>>
-FusedOperator::FusedOpIntervals(
+NgraphOperator::NgraphOpIntervals(
     std::vector<std::unique_ptr<paddle::framework::OperatorBase>>* ops) {
   std::vector<std::vector<std::vector<std::unique_ptr<OperatorBase>>::iterator>>
       intervals;
@@ -185,7 +184,7 @@ FusedOperator::FusedOpIntervals(
   return intervals;
 }
 
-FusedOperator::FusedOperator(
+NgraphOperator::NgraphOperator(
     const ProgramDesc& prog, size_t block_id,
     std::vector<std::unique_ptr<OperatorBase>>::iterator start,
     std::vector<std::unique_ptr<OperatorBase>>::iterator end,
@@ -215,7 +214,7 @@ FusedOperator::FusedOperator(
   Process();
 }
 
-void FusedOperator::Process() {
+void NgraphOperator::Process() {
   auto& bdesc = pdesc_.Block(block_);
   for (auto& var : bdesc.AllVars()) {
     if (!(var->GetType() == proto::VarType::SELECTED_ROWS ||
@@ -251,8 +250,8 @@ void FusedOperator::Process() {
   }
 }
 
-void FusedOperator::RunImpl(const Scope& scope,
-                            const platform::Place& place) const {
+void NgraphOperator::RunImpl(const Scope& scope,
+                             const platform::Place& place) const {
   op_state ng_op_state = PARTIAL_TEST;
   auto& bdesc = pdesc_.Block(block_);
   for (auto* op : bdesc.AllOps()) {
@@ -266,19 +265,19 @@ void FusedOperator::RunImpl(const Scope& scope,
     ng_op_state = ng_op_state == PARTIAL_TEST ? FULL_TEST : FULL_TRAIN;
   }
 
-  NgraphOperator ngraph_op(scope, place, fused_ops_, var_type_map_,
-                           persistables_, fetches_, post_op_inputs_,
-                           ng_op_state);
-  ngraph_op.Run(scope, place);
+  NgraphEngine ngraph_engine(scope, place, fused_ops_, var_type_map_,
+                             persistables_, fetches_, post_op_inputs_,
+                             ng_op_state);
+  ngraph_engine.Run(scope, place);
 }
 
 std::unordered_map<std::string, std::shared_ptr<ngraph::Function>>
-    NgraphOperator::func_cache_ = {};
+    NgraphEngine::func_cache_ = {};
 
-std::shared_ptr<ngraph::runtime::Backend> NgraphOperator::backend_ =
+std::shared_ptr<ngraph::runtime::Backend> NgraphEngine::backend_ =
     ngraph::runtime::Backend::create("CPU");
 
-void NgraphOperator::GetNgInputShape(std::shared_ptr<OperatorBase> op) {
+void NgraphEngine::GetNgInputShape(std::shared_ptr<OperatorBase> op) {
   op->RuntimeInferShape(scope_, place_);
   for (auto& var_name_item : op->Inputs()) {
     for (auto& var_name : var_name_item.second) {
@@ -301,7 +300,7 @@ void NgraphOperator::GetNgInputShape(std::shared_ptr<OperatorBase> op) {
   }
 }
 
-void NgraphOperator::BuildNgNodes() {
+void NgraphEngine::BuildNgNodes() {
   for (auto& var_name : var_out_) {
     if (var_node_map_->find(var_name) == var_node_map_->end()) {
       auto* var = scope_.FindVar(var_name);
@@ -323,7 +322,7 @@ void NgraphOperator::BuildNgNodes() {
   }
 }
 
-void NgraphOperator::BuildNgIO() {
+void NgraphEngine::BuildNgIO() {
   std::unordered_set<std::string> inputs;
   std::unordered_set<std::string> outputs;
 
@@ -395,7 +394,7 @@ void NgraphOperator::BuildNgIO() {
   }
 }
 
-void NgraphOperator::BuildNgFunction() {
+void NgraphEngine::BuildNgFunction() {
   BuildNgNodes();
   ngraph_function_ = nullptr;
   ngraph::NodeVector func_outputs;
@@ -416,7 +415,7 @@ void NgraphOperator::BuildNgFunction() {
       std::make_shared<ngraph::Function>(func_outputs, func_inputs);
 }
 
-std::shared_ptr<std::string> NgraphOperator::GetCacheKey() {
+std::shared_ptr<std::string> NgraphEngine::GetCacheKey() {
   auto cache_key = std::make_shared<std::string>("");
   *cache_key += std::to_string(fused_ops_.size());
   for (auto& op : fused_ops_) {
@@ -444,7 +443,7 @@ std::shared_ptr<std::string> NgraphOperator::GetCacheKey() {
   return cache_key;
 }
 
-void NgraphOperator::GetNgFunction() {
+void NgraphEngine::GetNgFunction() {
   bool cache_on = true;
   if (cache_on) {
     std::string cache_key_val = *GetCacheKey();
@@ -459,8 +458,7 @@ void NgraphOperator::GetNgFunction() {
   }
 }
 
-void NgraphOperator::Run(const Scope& scope,
-                         const platform::Place& place) const {
+void NgraphEngine::Run(const Scope& scope, const platform::Place& place) const {
   std::vector<std::shared_ptr<ngraph::runtime::Tensor>> t_in;
   std::vector<std::shared_ptr<ngraph::runtime::Tensor>> t_out;
 
@@ -545,7 +543,6 @@ void NgraphOperator::Run(const Scope& scope,
   }
 
   backend_->call(ngraph_function_, t_out, t_in);
-}  // NgraphOperator::RunImpl
+}  // NgraphEngine::RunImpl
 }  // namespace framework
 }  // namespace paddle
-#endif

paddle/fluid/framework/ngraph_operator.h

@@ -14,8 +14,6 @@ limitations under the License. */
 
 #pragma once
 
-#ifdef PADDLE_WITH_NGRAPH
-
 #include <algorithm>
 #include <string>
 #include <unordered_map>
@@ -34,14 +32,14 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-class FusedOperator : public OperatorBase {
+class NgraphOperator : public OperatorBase {
  public:
   static std::vector<
       std::vector<std::vector<std::unique_ptr<OperatorBase>>::iterator>>
-  FusedOpIntervals(
+  NgraphOpIntervals(
       std::vector<std::unique_ptr<paddle::framework::OperatorBase>>* ops);
 
-  explicit FusedOperator(
+  explicit NgraphOperator(
       const ProgramDesc& prog, size_t block_id,
       std::vector<std::unique_ptr<OperatorBase>>::iterator start,
       std::vector<std::unique_ptr<OperatorBase>>::iterator end,
@@ -64,4 +62,3 @@ class FusedOperator : public OperatorBase {
 };
 }  // namespace framework
 }  // namespace paddle
-#endif

paddle/fluid/framework/op_desc.cc

@@ -239,6 +239,23 @@ void OpDesc::SetOutput(const std::string &param_name,
   this->outputs_[param_name] = args;
 }
 
+bool OpDesc::HasProtoAttr(const std::string &name) const {
+  auto &op_info = OpInfoMap::Instance();
+  if (op_info.Has(desc_.type())) {
+    auto op_info_ptr = op_info.Get(desc_.type());
+    if (op_info_ptr.HasOpProtoAndChecker()) {
+      const proto::OpProto &proto = op_info_ptr.Proto();
+      for (int i = 0; i != proto.attrs_size(); ++i) {
+        const proto::OpProto::Attr &attr = proto.attrs(i);
+        if (attr.name() == name) {
+          return true;
+        }
+      }
+    }
+  }
+  return false;
+}
+
 proto::AttrType OpDesc::GetAttrType(const std::string &name) const {
   auto it = attrs_.find(name);
   PADDLE_ENFORCE(it != attrs_.end(), "Attribute %s is not found", name);

paddle/fluid/framework/op_desc.h

@@ -65,6 +65,8 @@ class OpDesc {
     return attrs_.find(name) != attrs_.end();
   }
 
+  bool HasProtoAttr(const std::string &name) const;
+
   proto::AttrType GetAttrType(const std::string &name) const;
 
   std::vector<std::string> AttrNames() const;

paddle/fluid/framework/parallel_executor.cc

@@ -30,13 +30,36 @@ limitations under the License. */
 #include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h"
 #include "paddle/fluid/platform/profiler.h"
 
+#ifdef WITH_GPERFTOOLS
+#include "gperftools/profiler.h"
+#endif
+DEFINE_string(pe_profile_fname, "",
+              "Profiler filename for PE, which generated by gperftools."
+              "Only valid when compiled `WITH_PRIFILER=ON`. Empty if disable.");
+
 namespace paddle {
 namespace framework {
 
+static std::once_flag gProfileOnce;
+#ifdef WITH_GPERFTOOLS
+static bool gProfileStarted = false;
+#endif
 class ParallelExecutorPrivate {
  public:
   explicit ParallelExecutorPrivate(const std::vector<platform::Place> &places)
-      : places_(places) {}
+      : places_(places) {
+    if (!FLAGS_pe_profile_fname.empty()) {
+      std::call_once(gProfileOnce, [] {
+#ifdef WITH_GPERFTOOLS
+        ProfilerStart(FLAGS_pe_profile_fname.c_str());
+        gProfileStarted = true;
+#else
+        LOG(WARNING) << "Paddle is not compiled with gperftools. "
+                        "FLAGS_pe_profile_fname will be ignored";
+#endif
+      });
+    }
+  }
 
   ~ParallelExecutorPrivate() {
     if (own_local_scope_) {
@@ -270,6 +293,12 @@ void ParallelExecutor::BCastParamsToDevices(
 
 void ParallelExecutor::Run(const std::vector<std::string> &fetch_tensors,
                            const std::string &fetched_var_name) {
+#ifdef WITH_GPERFTOOLS
+  if (gProfileStarted) {
+    ProfilerFlush();
+  }
+#endif
+
   platform::RecordBlock b(0);
 #ifdef PADDLE_WITH_CUDA
   if (!gcs_.empty()) {

paddle/fluid/imperative/CMakeLists.txt

+cc_library(layer SRCS layer.cc DEPS proto_desc operator)
+cc_library(tracer SRCS tracer.cc DEPS proto_desc)
+cc_library(engine SRCS engine.cc)

paddle/fluid/imperative/engine.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/imperative/engine.h"
+
+#include <mutex>  // NOLINT
+#include <vector>
+
+#include "glog/logging.h"
+
+namespace paddle {
+namespace imperative {
+
+static std::once_flag init_engine;
+static Engine* engine;
+
+class DummyEngine : public Engine {
+ public:
+  void Enqueue(Runnable* runnable) override {
+    queued_runnables_.push_back(runnable);
+  }
+
+  size_t Size() const override { return queued_runnables_.size(); }
+
+  void Sync() override {
+    for (Runnable* l : queued_runnables_) {
+      LOG(INFO) << "running " << reinterpret_cast<void*>(l);
+    }
+    queued_runnables_.clear();
+  }
+
+ private:
+  std::vector<Runnable*> queued_runnables_;
+};
+
+Engine* GetEngine() {
+  std::call_once(init_engine, []() { engine = new DummyEngine(); });
+  return engine;
+}
+
+}  // namespace imperative
+}  // namespace paddle

paddle/fluid/imperative/engine.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.
+
+#pragma once
+
+#include <cstddef>
+#include <cstdint>
+
+namespace paddle {
+namespace imperative {
+
+struct Runnable {};
+
+class Engine {
+ public:
+  virtual ~Engine() {}
+
+  virtual void Enqueue(Runnable* runnable) = 0;
+
+  virtual size_t Size() const = 0;
+
+  virtual void Sync() = 0;
+};
+
+Engine* GetEngine();
+
+}  // namespace imperative
+}  // namespace paddle

paddle/fluid/imperative/layer.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/imperative/layer.h"
+#include <deque>
+#include <limits>
+#include <map>
+#include <random>
+#include <utility>
+
+#include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/string/printf.h"
+
+namespace paddle {
+namespace imperative {
+
+using framework::Variable;
+
+void AddTo(Variable* src, Variable* dst) {
+  framework::LoDTensor* dst_tensor = dst->GetMutable<framework::LoDTensor>();
+  framework::LoDTensor* src_tensor = src->GetMutable<framework::LoDTensor>();
+  PADDLE_ENFORCE(dst_tensor->numel() == src_tensor->numel(), "%lld vs %lld",
+                 dst_tensor->numel(), src_tensor->numel());
+  float* dst_data = dst_tensor->mutable_data<float>(platform::CPUPlace());
+  const float* src_data = src_tensor->data<float>();
+  for (size_t i = 0; i < src_tensor->numel(); ++i) {
+    dst_data[i] += src_data[i];
+  }
+}
+
+class Autograd {
+ public:
+  explicit Autograd(framework::Scope* scope) : scope_(scope) {}
+
+  void RunBackward(VarBase* var) {
+    PADDLE_ENFORCE(var->pre_op_->op_desc_);
+    // TODO(panyx0718): Only create for vars that "require_grad"
+    (*var->pre_op_->output_vars_)[var->pre_op_out_idx_]->grads_ = var->grads_;
+
+    std::deque<OpBase*> ready;
+    ready.push_back(var->pre_op_);
+
+    std::map<OpBase*, int> dep_counts = ComputeDepCounts(var->pre_op_);
+
+    while (!ready.empty()) {
+      OpBase* ready_op = ready.front();
+      ready.pop_front();
+      std::vector<Variable*> input_grads = ready_op->ApplyGrad(scope_);
+
+      for (size_t i = 0; i < input_grads.size(); ++i) {
+        if (!input_grads[i]) continue;
+        OpBase* pre_op = ready_op->pre_ops_->at(i);
+        if (!pre_op) continue;
+
+        dep_counts[pre_op] -= 1;
+        PADDLE_ENFORCE(dep_counts[pre_op] >= 0);
+        bool pre_op_ready = dep_counts[pre_op] == 0;
+        if (pre_op_ready) {
+          ready.push_back(pre_op);
+        }
+      }
+    }
+  }
+
+ private:
+  std::map<OpBase*, int> ComputeDepCounts(OpBase* op) {
+    std::map<OpBase*, int> ret;
+
+    std::deque<OpBase*> queue;
+    queue.push_back(op);
+    std::unordered_set<OpBase*> visited;
+    visited.insert(op);
+    while (!queue.empty()) {
+      OpBase* candidate = queue.front();
+      queue.pop_front();
+      for (OpBase* pre_op : *(candidate->pre_ops_)) {
+        if (!pre_op) continue;
+        if (visited.find(pre_op) == visited.end()) {
+          visited.insert(pre_op);
+          queue.push_back(pre_op);
+        }
+        ret[pre_op] += 1;
+      }
+    }
+
+    return ret;
+  }
+
+  framework::Scope* scope_;
+};
+
+framework::Variable* CreateVariable(const std::string& name,
+                                    const framework::DDim& dim, float val,
+                                    framework::Scope* scope,
+                                    bool random_name = true) {
+  std::string varname = name;
+  if (random_name) {
+    std::mt19937 rng;
+    rng.seed(std::random_device()());
+    std::uniform_int_distribution<std::mt19937::result_type> dist6(
+        1, std::numeric_limits<int>::max());
+    int id = dist6(rng);
+    varname = string::Sprintf("%s@%d", varname, id);
+  }
+
+  VLOG(3) << "creating var " << varname;
+  framework::Variable* var = scope->Var(varname);
+  framework::LoDTensor* tensor = var->GetMutable<framework::LoDTensor>();
+
+  float* data = tensor->mutable_data<float>(dim, platform::CPUPlace());
+  std::fill(data, data + tensor->numel(), val);
+  return var;
+}
+
+framework::LoDTensor& VarBase::Grad() {
+  VLOG(3) << "get var grad " << var_desc_->Name();
+  return *grads_->GetMutable<framework::LoDTensor>();
+}
+
+void VarBase::ApplyGrad(framework::Scope* scope, Variable* grad) {
+  VLOG(3) << "apply var grad " << var_desc_->Name() << " "
+          << grad->Get<framework::LoDTensor>().data<float>()[0];
+  if (!grads_) {
+    grads_ =
+        CreateVariable(string::Sprintf("%s@IGrad", var_desc_->Name()),
+                       var_->Get<framework::LoDTensor>().dims(), 0.0, scope);
+  }
+  AddTo(grad, grads_);
+  VLOG(3) << "grad_ after apply var grad " << var_desc_->Name() << " "
+          << grads_->Get<framework::LoDTensor>().data<float>()[0];
+}
+
+std::vector<Variable*> OpBase::ApplyGrad(framework::Scope* scope) {
+  VLOG(3) << "op grad " << grad_op_desc_->Type();
+
+  for (const std::string& grad_invar : grad_op_desc_->InputArgumentNames()) {
+    if (grad_to_var_->find(grad_invar) == grad_to_var_->end()) {
+      // grad op inputs can be forward inputs, so not in grad_to_var.
+      continue;
+    }
+    VLOG(3) << "op grad in var " << grad_invar;
+    block_->FindRecursiveOrCreateVar(grad_invar);
+    framework::Variable* var = scope->Var(grad_invar);
+    const std::string& invar = grad_to_var_->at(grad_invar);
+    for (VarBase* varbase : *output_vars_) {
+      // Use the accumulated grads_ by sharing the input with grads_.
+      if (varbase->var_desc_->Name() == invar) {
+        var->GetMutable<framework::LoDTensor>()->ShareDataWith(
+            varbase->grads_->Get<framework::LoDTensor>());
+        break;
+      }
+    }
+  }
+
+  for (const std::string& outvar : grad_op_desc_->OutputArgumentNames()) {
+    VLOG(3) << "grad outvar " << outvar;
+    block_->FindRecursiveOrCreateVar(outvar);
+    framework::Variable* var = scope->Var(outvar);
+    if (!var->IsInitialized()) {
+      framework::VarDesc* var_desc = block_->FindVar(outvar);
+      if (var_desc->GetType() == framework::proto::VarType::LOD_TENSOR) {
+        var->GetMutable<framework::LoDTensor>();
+      } else {
+        LOG(ERROR) << "tracer doesn't support yet";
+      }
+    }
+  }
+  grad_op_desc_->InferShape(*block_);
+  grad_op_desc_->InferVarType(block_);
+  std::unique_ptr<framework::OperatorBase> opbase =
+      framework::OpRegistry::CreateOp(*grad_op_desc_);
+
+  opbase->Run(*scope, platform::CPUPlace());
+
+  // `ret` matches exactly with `input_vars_` of forward op.
+  std::vector<Variable*> ret;
+  for (size_t i = 0; i < input_vars_->size(); ++i) {
+    bool found = false;
+    for (const std::string& outvar : grad_op_desc_->OutputArgumentNames()) {
+      Variable* var = scope->FindVar(outvar);
+      VarBase* origin_var = (*input_vars_)[i];
+      std::string orig_var = grad_to_var_->at(outvar);
+      PADDLE_ENFORCE(origin_var->var_desc_->Name() == orig_var);
+      VLOG(3) << "apply grad " << outvar << " with origin " << orig_var;
+      origin_var->ApplyGrad(scope, var);
+      found = true;
+      ret.push_back(var);
+      // TODO(panyx0718): There might be another outvar with the same name.
+      // In that case, it doesn't matter the first one or the second one is
+      // used.
+      break;
+    }
+    if (!found) {
+      ret.push_back(nullptr);
+    }
+  }
+  return ret;
+}
+
+void VarBase::RunBackward(framework::Scope* scope) {
+  grads_ = CreateVariable(framework::GradVarName(var_desc_->Name()),
+                          var_->Get<framework::LoDTensor>().dims(), 1.0, scope,
+                          false);
+  if (!pre_op_) return;
+  Autograd(scope).RunBackward(this);
+}
+
+}  // namespace imperative
+}  // namespace paddle

paddle/fluid/imperative/layer.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.
+
+#pragma once
+
+#include <string>
+#include <vector>
+#include "paddle/fluid/framework/op_desc.h"
+#include "paddle/fluid/framework/operator.h"
+#include "paddle/fluid/framework/scope.h"
+#include "paddle/fluid/framework/var_desc.h"
+#include "paddle/fluid/platform/enforce.h"
+
+namespace paddle {
+namespace imperative {
+
+class OpBase;
+
+class VarBase {
+ public:
+  VarBase()
+      : pre_op_(nullptr),
+        pre_op_out_idx_(-1),
+        var_desc_(nullptr),
+        var_(nullptr),
+        grads_(nullptr) {}
+
+  virtual ~VarBase() {}
+
+  void ApplyGrad(framework::Scope* scope, framework::Variable* grad);
+
+  void RunBackward(framework::Scope* scope);
+
+  framework::LoDTensor& Grad();
+
+  OpBase* pre_op_;
+  int pre_op_out_idx_;
+
+  framework::VarDesc* var_desc_;
+  framework::Variable* var_;
+  framework::Variable* grads_;
+};
+
+class OpBase {
+ public:
+  OpBase()
+      : input_vars_(new std::vector<VarBase*>()),
+        output_vars_(new std::vector<VarBase*>()),
+        pre_ops_(new std::vector<OpBase*>()),
+        pre_ops_out_idx_(new std::vector<int>()),
+        op_desc_(nullptr),
+        grad_op_desc_(nullptr) {}
+
+  virtual ~OpBase() {
+    delete input_vars_;
+    delete output_vars_;
+
+    delete pre_ops_;
+    delete pre_ops_out_idx_;
+
+    if (grad_op_desc_) delete grad_op_desc_;
+    if (grad_to_var_) delete grad_to_var_;
+  }
+
+  std::vector<framework::Variable*> ApplyGrad(framework::Scope* scope);
+
+  std::vector<VarBase*>* input_vars_;
+  std::vector<VarBase*>* output_vars_;
+  std::vector<OpBase*>* pre_ops_;
+  std::vector<int>* pre_ops_out_idx_;
+  framework::OpDesc* op_desc_;
+
+  framework::OpDesc* grad_op_desc_;
+  std::unordered_map<std::string, std::string>* grad_to_var_;
+  framework::BlockDesc* block_;
+};
+
+class Layer {
+ public:
+  virtual ~Layer() {}
+
+  virtual std::vector<VarBase> Forward(const std::vector<VarBase>& inputs) {
+    std::vector<VarBase> vars;
+    return vars;
+  }
+
+  virtual void Backward() { LOG(ERROR) << "To support customize"; }
+};
+
+}  // namespace imperative
+}  // namespace paddle

paddle/fluid/imperative/tracer.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/imperative/tracer.h"
+
+namespace paddle {
+namespace imperative {}  // namespace imperative
+}  // namespace paddle

paddle/fluid/imperative/tracer.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.
+
+#pragma once
+
+#include <map>
+#include <string>
+#include <vector>
+
+#include "paddle/fluid/framework/op_desc.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/scope.h"
+#include "paddle/fluid/imperative/engine.h"
+#include "paddle/fluid/imperative/layer.h"
+
+namespace paddle {
+namespace imperative {
+
+void CreateGradOp(const framework::OpDesc& op_desc,
+                  const std::unordered_set<std::string>& no_grad_set,
+                  const std::vector<framework::BlockDesc*>& grad_sub_block,
+                  framework::OpDesc** grad_op_desc,
+                  std::unordered_map<std::string, std::string>* grad_to_var) {
+  std::vector<std::unique_ptr<framework::OpDesc>> grad_op_descs =
+      framework::OpInfoMap::Instance()
+          .Get(op_desc.Type())
+          .GradOpMaker()(op_desc, no_grad_set, grad_to_var, grad_sub_block);
+  PADDLE_ENFORCE(grad_op_descs.size() == 1, "Only support 1 grad op now.");
+  // TODO(panyx0718): Leak?
+  *grad_op_desc = grad_op_descs[0].release();
+}
+
+class Tracer {
+ public:
+  explicit Tracer(framework::BlockDesc* root_block) : root_block_(root_block) {
+    root_scope_ = new framework::Scope();
+    scopes_[root_block_] = root_scope_;
+  }
+
+  virtual ~Tracer() { delete root_scope_; }
+
+  void Trace(OpBase* op, const std::vector<VarBase*>& inputs,
+             const std::vector<VarBase*>& outputs,
+             framework::BlockDesc* block) {
+    framework::Scope* scope = GetScope(block);
+    framework::OpDesc* op_desc = op->op_desc_;
+    VLOG(3) << "tracer tracing " << op_desc->Type();
+    op_desc->InferShape(*block);
+    op_desc->InferVarType(block);
+    std::unique_ptr<framework::OperatorBase> op_base =
+        framework::OpRegistry::CreateOp(*op_desc);
+
+    *op->input_vars_ = inputs;
+    for (VarBase* input : inputs) {
+      const std::string vname = input->var_desc_->Name();
+      framework::Variable* var = scope->Var(vname);
+      input->var_ = var;
+      if (!var->IsInitialized()) {
+        framework::VarDesc* var_desc = block->FindVar(vname);
+        if (var_desc->GetType() == framework::proto::VarType::LOD_TENSOR) {
+          var->GetMutable<framework::LoDTensor>();
+        } else {
+          LOG(ERROR) << "tracer doesn't support yet";
+        }
+      }
+      if (input->pre_op_) {
+        op->pre_ops_->push_back(input->pre_op_);
+        op->pre_ops_out_idx_->push_back(input->pre_op_out_idx_);
+      } else {
+        op->pre_ops_->push_back(nullptr);
+      }
+    }
+
+    *op->output_vars_ = outputs;
+    for (size_t i = 0; i < outputs.size(); ++i) {
+      const std::string vname = outputs[i]->var_desc_->Name();
+      framework::Variable* var = scope->Var(vname);
+      if (!var->IsInitialized()) {
+        framework::VarDesc* var_desc = block->FindVar(vname);
+        if (var_desc->GetType() == framework::proto::VarType::LOD_TENSOR) {
+          var->GetMutable<framework::LoDTensor>();
+        } else {
+          LOG(ERROR) << "tracer doesn't support yet";
+        }
+      }
+      outputs[i]->var_ = var;
+      outputs[i]->pre_op_ = op;
+      outputs[i]->pre_op_out_idx_ = i;
+    }
+    op_base->Run(*scope, platform::CPUPlace());
+    framework::OpDesc* grad_op_desc;
+    auto grad_to_var = new std::unordered_map<std::string, std::string>();
+    CreateGradOp(*op_desc, {}, {block}, &grad_op_desc, grad_to_var);
+    op->grad_op_desc_ = grad_op_desc;
+    op->grad_to_var_ = grad_to_var;
+    op->block_ = block;
+  }
+
+  framework::Scope* GetScope(framework::BlockDesc* block) {
+    if (scopes_.find(block) != scopes_.end()) {
+      return scopes_.at(block);
+    }
+    framework::BlockDesc* parent_block = block->ParentBlock();
+    PADDLE_ENFORCE(scopes_.find(parent_block) != scopes_.end());
+    framework::Scope* scope = &scopes_[parent_block]->NewScope();
+    scopes_[block] = scope;
+    return scope;
+  }
+
+ private:
+  std::map<framework::BlockDesc*, framework::Scope*> scopes_;
+  framework::BlockDesc* root_block_;
+  framework::Scope* root_scope_;
+};
+
+}  // namespace imperative
+}  // namespace paddle

paddle/fluid/inference/analysis/passes/ir_graph_build_pass.cc

@@ -44,9 +44,10 @@ void IrGraphBuildPass::RunImpl(Argument *argument) {
     argument->SetMainProgram(program.release());
   } else if (argument->model_program_path_valid() &&
              argument->model_params_path_valid()) {
-    auto program =
-        LoadModel(argument->model_program_path(), argument->model_params_path(),
-                  argument->scope_ptr(), place, argument->model_from_memory());
+    auto program = LoadModel(
+        argument->model_program_path(), argument->model_params_path(),
+        argument->scope_ptr(), place,
+        argument->model_from_memory_valid() && argument->model_from_memory());
     argument->SetMainProgram(program.release());
   } else {
     PADDLE_THROW(

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

@@ -109,8 +109,12 @@ class Pool2dOpConverter : public OpConverter {
     }
 
     if (pool_type == "max") {
-      nvinfer1::DimsHW pre_pad(paddings[0], paddings[1]);
-      nvinfer1::DimsHW post_pad(paddings[0], paddings[1]);
+      // Under ceil mode, the pre_pad and post_pad are used to
+      // record the the padding size. In some ceil mode cases,
+      // we do not need padding, so we initialize the two vars to 0.
+
+      nvinfer1::DimsHW pre_pad(0, 0);
+      nvinfer1::DimsHW post_pad(0, 0);
       if (ceil_mode) {
         // If ceil mode is true, we will pad the appropriate size to the input.
         DealCeilMode(input_shape, ksize, strides, paddings, &pre_pad, &post_pad,

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

-set(INFERENCE_EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis_predictor)
+set(INFERENCE_EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis_predictor benchmark)
 
 if(WITH_GPU AND TENSORRT_FOUND)
     set(INFERENCE_EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} analysis ${analysis_deps} ir_pass_manager analysis_predictor)

paddle/fluid/inference/tests/api/tester_helper.h

@@ -30,8 +30,10 @@
 #include "paddle/fluid/inference/api/helper.h"
 #include "paddle/fluid/inference/tests/api/config_printer.h"
 #include "paddle/fluid/inference/tests/test_helper.h"
+#include "paddle/fluid/inference/utils/benchmark.h"
 #include "paddle/fluid/platform/profiler.h"
 
+DEFINE_string(model_name, "", "model name");
 DEFINE_string(infer_model, "", "model path");
 DEFINE_string(infer_data, "", "data file");
 DEFINE_int32(batch_size, 1, "batch size.");
@@ -40,6 +42,8 @@ DEFINE_bool(test_all_data, false, "Test the all dataset in data file.");
 DEFINE_int32(num_threads, 1, "Running the inference program in multi-threads.");
 DEFINE_bool(use_analysis, true,
             "Running the inference program in analysis mode.");
+DEFINE_bool(record_benchmark, false,
+            "Record benchmark after profiling the model");
 
 DECLARE_bool(profile);
 DECLARE_int32(paddle_num_threads);
@@ -192,8 +196,16 @@ void TestOneThreadPrediction(
         predictor->Run(inputs[j], outputs, batch_size);
       }
     }
-    PrintTime(batch_size, num_times, 1, 0, run_timer.toc() / num_times,
-              inputs.size());
+
+    double latency = run_timer.toc() / num_times;
+    PrintTime(batch_size, num_times, 1, 0, latency, inputs.size());
+    if (FLAGS_record_benchmark) {
+      Benchmark benchmark;
+      benchmark.SetName(FLAGS_model_name);
+      benchmark.SetBatchSize(batch_size);
+      benchmark.SetLatency(latency);
+      benchmark.PersistToFile("benchmark_record.txt");
+    }
   }
 }
 

paddle/fluid/inference/tests/api/trt_models_tester.cc

@@ -135,6 +135,9 @@ TEST(TensorRT_resnext50, compare) {
 
 TEST(TensorRT_resnext50, profile) {
   std::string model_dir = FLAGS_infer_model + "/resnext50";
+  // Set FLAGS_record_benchmark to true to record benchmark to file.
+  // FLAGS_record_benchmark=true;
+  FLAGS_model_name = "resnext50";
   profile(model_dir, /* use_analysis */ true, FLAGS_use_tensorrt);
 }
 

paddle/fluid/inference/utils/benchmark.cc

@@ -30,7 +30,7 @@ std::string Benchmark::SerializeToString() const {
   ss << '\n';
 
   ss << name_ << "\t";
-  ss << batch_size_ << "\t";
+  ss << batch_size_ << "\t\t";
   ss << num_threads_ << "\t";
   ss << latency_ << "\t";
   ss << 1000.0 / latency_;

paddle/fluid/inference/utils/visualizer.cc

@@ -26,9 +26,6 @@ DEFINE_string(model_dir, "", "model directory");
 DEFINE_string(model_program_path, "", "model program path");
 DEFINE_string(model_params_path, "", "model params path");
 
-USE_PASS(graph_viz_pass);
-USE_PASS(graph_to_program_pass);
-
 using paddle::inference::analysis::Argument;
 
 namespace paddle {
@@ -40,7 +37,6 @@ void Visualizer::SetArgument(Argument *argument) { argument_ = argument; }
 bool Visualizer::Run() {
   paddle::framework::InitDevices(false);
   paddle::inference::analysis::Analyzer().Run(argument_);
-
   return true;
 }
 
@@ -77,7 +73,7 @@ int main(int argc, char *argv[]) {
 
   // Only 1 pass, default filename is 0_ir_origin.dot
   // For more details, looking for paddle::inference::analysis::IRPassManager
-  argument.SetIrAnalysisPasses({"graph_viz_pass"});
+  argument.SetIrAnalysisPasses({"infer_clean_graph_pass", "graph_viz_pass"});
 
   std::unique_ptr<paddle::framework::Scope> scope{
       new paddle::framework::Scope()};
@@ -90,3 +86,7 @@ int main(int argc, char *argv[]) {
 
   return 0;
 }
+
+USE_PASS(infer_clean_graph_pass);
+USE_PASS(graph_viz_pass);
+USE_PASS(graph_to_program_pass);

paddle/fluid/operators/activation_op.h

@@ -301,23 +301,22 @@ template <typename T>
 struct GeluFunctor : public BaseActivationFunctor<T> {
   template <typename Device, typename X, typename Out>
   void operator()(Device d, X x, Out out) const {
-    auto temp =
-        ((x * static_cast<T>(M_SQRT1_2)).erf()).template cast<T>().eval();
+    auto temp = (x * static_cast<T>(M_SQRT1_2)).erf();
     out.device(d) = x * static_cast<T>(0.5) * (static_cast<T>(1) + temp);
   }
 };
 
 template <typename T>
 struct GeluGradFunctor : BaseActivationFunctor<T> {
-  bool Inplace() const { return IsInplace("gelu"); }
   template <typename Device, typename X, typename Out, typename dOut,
             typename dX>
   void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
-    auto temp = (static_cast<T>(0.5 * M_2_SQRTPI * M_SQRT1_2) * x *
-                 ((-static_cast<T>(0.5) * x.square()).exp()))
-                    .template cast<T>()
-                    .eval();
-    dx.device(d) = dout * (out / x + temp);
+    auto first = static_cast<T>(0.5) *
+                 (static_cast<T>(1) + ((x * static_cast<T>(M_SQRT1_2)).erf()));
+
+    auto second = static_cast<T>(0.5 * M_2_SQRTPI * M_SQRT1_2) * x *
+                  (-static_cast<T>(0.5) * x.square()).exp();
+    dx.device(d) = dout * (first + second);
   }
 };
 

paddle/fluid/operators/bilinear_tensor_product_op.cu

@@ -12,7 +12,6 @@ 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. */
 
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/bilinear_tensor_product_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/bpr_loss_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/fluid/operators/bpr_loss_op.h"
+
+namespace paddle {
+namespace operators {
+
+class BprLossOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
+    PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Y"), "Output(Y) should be not null.");
+
+    auto x_dims = ctx->GetInputDim("X");
+    auto label_dims = ctx->GetInputDim("Label");
+    int rank = x_dims.size();
+    PADDLE_ENFORCE_EQ(rank, label_dims.size(),
+                      "Input(X) and Input(Label) shall have the same rank.");
+    PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 0, rank - 1),
+                      framework::slice_ddim(label_dims, 0, rank - 1),
+                      "Input(X) and Input(Label) shall have the same shape "
+                      "except the last dimension.");
+
+    auto y_dims = x_dims;
+    y_dims[rank - 1] = 1;
+    ctx->SetOutputDim("Y", y_dims);
+    ctx->ShareLoD("X", /*->*/ "Y");
+  }
+
+ protected:
+  // Explicitly set that the data type of computation kernel of Seq-bpr
+  // is determined by its input "X".
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(
+        framework::ToDataType(ctx.Input<Tensor>("X")->type()),
+        platform::CPUPlace());
+  }
+};
+
+class BprLossGradientOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
+    PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null.");
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Y")),
+                   "Input(Y@GRAD) shoudl be not null.");
+    PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")),
+                   "Output(X@GRAD) should be not null.");
+
+    auto x_dims = ctx->GetInputDim("X");
+    auto label_dims = ctx->GetInputDim("Label");
+    auto dy_dims = ctx->GetInputDim(framework::GradVarName("Y"));
+    int rank = x_dims.size();
+    PADDLE_ENFORCE_EQ(dy_dims.size(), rank,
+                      "Input(Y@Grad) and Input(X) should have the same rank.");
+    PADDLE_ENFORCE_EQ(label_dims.size(), rank,
+                      "Input(Label) and Input(X) should have the same rank.");
+    PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 0, rank - 1),
+                      framework::slice_ddim(label_dims, 0, rank - 1),
+                      "The Input(X) and Input(Label) should have the same "
+                      "shape except the last dimension.");
+    PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 0, rank - 1),
+                      framework::slice_ddim(dy_dims, 0, rank - 1),
+                      "The Input(X) and Input(Y@Grad) should have the same "
+                      "shape except the last dimension.");
+    PADDLE_ENFORCE_EQ(dy_dims[rank - 1], 1,
+                      "The last dimension of Input(Y@Grad) should be 1.");
+    PADDLE_ENFORCE_EQ(label_dims[rank - 1], 1,
+                      " the last dimension of Input(Label) should be 1.");
+    ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
+    ctx->ShareLoD("X", framework::GradVarName("X"));
+  }
+
+ protected:
+  // Explicitly set that the data type of computation kernel of cross_entropy
+  // is determined by its input "X".
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(
+        framework::ToDataType(ctx.Input<Tensor>("X")->type()),
+        platform::CPUPlace());
+  }
+};
+
+class BprLossOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X",
+             "(Tensor, default Tensor<float>), a tensor whose last dimension "
+             "size is equal to the number of classes. This input is a "
+             "real number.");
+    AddInput(
+        "Label",
+        "(Tensor), the tensor which represents the ground truth. It has the "
+        "same shape with 'X' except the last dimension. the last dimension "
+        "size is 1.");
+    AddOutput("Y",
+              "(Tensor, default Tensor<float>), a tensor whose shape is same "
+              "with 'X' except that the last dimension size is 1. It "
+              "represents the sequence bpr loss.");
+    AddComment(R"DOC(
+Bayesian Personalized Ranking Loss Operator.
+
+This operator belongs to pairwise ranking loss. Label is the desired item.
+The loss at a given point in one session is defined as:
+$Y[i] = -\frac{1}{N_{i}} * \sum_{j=0}^{N_{i}}\log(\sigma(X[i, Label[i]]-X[i, j]))$
+
+Learn more details by reading paper <session-based recommendations with recurrent
+neural networks>(https://arxiv.org/abs/1511.06939)
+
+)DOC");
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+using CPUCtx = paddle::platform::CPUDeviceContext;
+
+REGISTER_OPERATOR(bpr_loss, ops::BprLossOp, ops::BprLossOpMaker,
+                  paddle::framework::DefaultGradOpDescMaker<true>);
+REGISTER_OPERATOR(bpr_loss_grad, ops::BprLossGradientOp);
+REGISTER_OP_CPU_KERNEL(bpr_loss, ops::BprLossOpKernel<CPUCtx, float>,
+                       ops::BprLossOpKernel<CPUCtx, double>);
+REGISTER_OP_CPU_KERNEL(bpr_loss_grad,
+                       ops::BprLossGradientOpKernel<CPUCtx, float>,
+                       ops::BprLossGradientOpKernel<CPUCtx, double>);

paddle/fluid/operators/bpr_loss_op.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/fluid/framework/eigen.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/math/math_function.h"
+#include "paddle/fluid/platform/for_range.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+/*Todo:
+ *Find a way to adapt TolerableValue, using blas or eigen.
+ */
+template <typename T>
+struct TolerableValue {
+  HOSTDEVICE T operator()(const T& x) const {
+    PADDLE_ASSERT(std::is_floating_point<T>::value);
+    const T kApproInf = 1e20;
+    if (x == INFINITY) return kApproInf;
+    if (x == -INFINITY) return -kApproInf;
+    return x;
+  }
+};
+
+template <typename DeviceContext, typename T>
+class BprLossOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x = ctx.Input<Tensor>("X");
+    auto* label = ctx.Input<Tensor>("Label");
+    auto* y = ctx.Output<Tensor>("Y");
+    y->mutable_data<T>(ctx.GetPlace());
+    int rank = x->dims().size();
+
+    Tensor x_2d = framework::ReshapeToMatrix(*x, rank - 1);
+    Tensor labels_2d = framework::ReshapeToMatrix(*label, rank - 1);
+    Tensor y_2d = framework::ReshapeToMatrix(*y, rank - 1);
+
+    const framework::Tensor* logits = &x_2d;
+    const framework::Tensor* labels = &labels_2d;
+    framework::Tensor* out = &y_2d;
+
+    const int step_size = logits->dims()[0];
+    const int class_num = logits->dims()[1];
+    const T* logits_data = logits->data<T>();
+    T* loss_data = out->data<T>();
+
+    const int64_t* label_data = labels->data<int64_t>();
+    for (int i = 0; i < step_size; ++i) {
+      int lbl_pos = label_data[i];
+      PADDLE_ENFORCE_GE(lbl_pos, 0);
+      PADDLE_ENFORCE_LT(lbl_pos, class_num);
+      int index_pos = i * class_num + lbl_pos;
+      T sum = static_cast<T>(0);
+      for (int j = 0; j < class_num; j++) {
+        if (j == lbl_pos) continue;
+        int index_neg = i * class_num + j;
+        sum += TolerableValue<T>()(-std::log(
+            1.0f + TolerableValue<T>()(std::exp(logits_data[index_neg] -
+                                                logits_data[index_pos]))));
+      }
+      loss_data[i] = -sum / (class_num - 1);
+    }
+  }
+};
+
+template <typename DeviceContext, typename T>
+class BprLossGradientOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x = ctx.Input<Tensor>("X");
+    auto* dy = ctx.Input<Tensor>(framework::GradVarName("Y"));
+    auto* label = ctx.Input<Tensor>("Label");
+    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+
+    const int step_size = x->dims()[0];
+    const int num_classes = x->dims()[1];
+    T* dx_data = dx->mutable_data<T>(ctx.GetPlace());
+    const T* dy_data = dy->data<T>();
+    const T* x_data = x->data<T>();
+    const int64_t* label_data = label->data<int64_t>();
+
+    for (size_t sample_id = 0; sample_id < step_size; sample_id++) {
+      for (size_t x_offset = sample_id * num_classes;
+           x_offset < (sample_id + 1) * num_classes; x_offset++) {
+        dx_data[x_offset] = static_cast<T>(0);
+      }
+      auto p_index = sample_id * num_classes + label_data[sample_id];
+      for (size_t ni = 0; ni < num_classes; ni++) {
+        if (label_data[sample_id] == ni) continue;
+        auto n_index = sample_id * num_classes + ni;
+        auto grad_ = -dy_data[sample_id] /
+                     ((num_classes - 1) *
+                      (1.0f + TolerableValue<T>()(std::exp(x_data[p_index] -
+                                                           x_data[n_index]))));
+        dx_data[p_index] += grad_;
+        dx_data[n_index] -= grad_;
+      }
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/concat_mkldnn_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 <memory>
+#include "paddle/fluid/operators/concat_op.h"
+#include "paddle/fluid/platform/mkldnn_helper.h"
+
+namespace paddle {
+namespace operators {
+
+using framework::DataLayout;
+using framework::Tensor;
+using mkldnn::memory;
+using mkldnn::primitive;
+using mkldnn::concat;
+using mkldnn::stream;
+using platform::to_void_cast;
+
+static void EnforceLayouts(const std::vector<const Tensor*> inputs) {
+  for (auto* input : inputs) {
+    const bool is_layout_correct = input->layout() == DataLayout::kMKLDNN;
+    const bool is_format_defined =
+        input->format() != memory::format::format_undef;
+    PADDLE_ENFORCE(is_layout_correct && is_format_defined,
+                   "Wrong layout/format set for Input tensor");
+  }
+}
+
+static memory::primitive_desc CreateMemPrimDesc(const Tensor& input,
+                                                const mkldnn::engine& engine) {
+  constexpr auto data_type = mkldnn::memory::f32;
+  const auto dims = paddle::framework::vectorize2int(input.dims());
+  const auto format = input.format();
+  auto description = memory::desc(dims, data_type, format);
+  auto mem_prim_desc = memory::primitive_desc(description, engine);
+  return mem_prim_desc;
+}
+
+static mkldnn::memory::format GetDstMemFormat(
+    const concat::primitive_desc& concat_pd) {
+  return (memory::format)concat_pd.dst_primitive_desc().desc().data.format;
+}
+
+static platform::CPUPlace GetCpuPlace(
+    const paddle::framework::ExecutionContext& ctx) {
+  auto place = ctx.GetPlace();
+  PADDLE_ENFORCE(paddle::platform::is_cpu_place(place),
+                 "It must use CPUPlace.");
+  return boost::get<platform::CPUPlace>(place);
+}
+
+static const mkldnn::engine& GetMKLDNNEngine(
+    const paddle::framework::ExecutionContext& ctx) {
+  auto& dev_ctx = ctx.template device_context<platform::MKLDNNDeviceContext>();
+  return dev_ctx.GetEngine();
+}
+
+template <typename T>
+class ConcatPrimitiveFactory {
+ public:
+  concat::primitive_desc CreateConcatPrimDescriptor(
+      const std::vector<const Tensor*> multi_input, Tensor* output,
+      int concat_axis, const mkldnn::engine& mkldnn_engine) {
+    CreateSourcesDescriptors(multi_input, mkldnn_engine);
+    auto dst_desc = CreateDstMemDescriptor(output);
+    return concat::primitive_desc(dst_desc, concat_axis, srcs_pd);
+  }
+
+  concat CreateConcatPrimitive(const concat::primitive_desc& concat_pd,
+                               Tensor* output, platform::CPUPlace place) {
+    CreateSourcePrimitiveAts();
+    dst_mem = CreateDstMemory(concat_pd, output, place);
+    return concat(concat_pd, inputs, dst_mem.get());
+  }
+
+ private:
+  memory::desc CreateDstMemDescriptor(Tensor* output) {
+    auto dst_dims = paddle::framework::vectorize2int(output->dims());
+    return memory::desc(dst_dims, platform::MKLDNNGetDataType<T>(),
+                        memory::format::any);
+  }
+
+  mkldnn::memory CreateDstMemory(const concat::primitive_desc& concat_pd,
+                                 Tensor* output, platform::CPUPlace place) {
+    return memory(concat_pd.dst_primitive_desc(),
+                  output->mutable_data<T>(place));
+  }
+
+  void CreateSourcesDescriptors(const std::vector<const Tensor*> multi_input,
+                                const mkldnn::engine& mkldnn_engine) {
+    for (size_t i = 0; i < multi_input.size(); i++) {
+      auto mem_prim_desc = CreateMemPrimDesc(*multi_input[i], mkldnn_engine);
+      srcs_pd.push_back(mem_prim_desc);
+      srcs.push_back(
+          memory(mem_prim_desc, to_void_cast(multi_input[i]->data<T>())));
+    }
+  }
+
+  void CreateSourcePrimitiveAts() {
+    inputs.reserve(srcs.size());
+    for (size_t i = 0; i < srcs.size(); i++) {
+      inputs.push_back(srcs[i]);
+    }
+  }
+
+ private:
+  std::vector<memory::primitive_desc> srcs_pd;
+  std::vector<memory> srcs;
+  std::vector<primitive::at> inputs;
+  boost::optional<memory> dst_mem;  // TODO(mgallus): change to std::optional
+};                                  // upon introduction of C++17 to paddle
+
+template <typename T>
+class ConcatMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
+ public:
+  void Compute(const paddle::framework::ExecutionContext& ctx) const override {
+    auto place = GetCpuPlace(ctx);
+    const auto& mkldnn_engine = GetMKLDNNEngine(ctx);
+
+    auto multi_input = ctx.MultiInput<Tensor>("X");
+    EnforceLayouts(multi_input);
+    Tensor* output = ctx.Output<Tensor>("Out");
+    int64_t concat_axis = static_cast<int64_t>(ctx.Attr<int>("axis"));
+
+    ConcatPrimitiveFactory<T> prim_creator;
+    auto concat_pd = prim_creator.CreateConcatPrimDescriptor(
+        multi_input, output, static_cast<int>(concat_axis), mkldnn_engine);
+    auto concat = prim_creator.CreateConcatPrimitive(concat_pd, output, place);
+    stream(stream::kind::eager).submit({concat}).wait();
+
+    output->set_layout(DataLayout::kMKLDNN);
+    output->set_format(GetDstMemFormat(concat_pd));
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_KERNEL(concat, MKLDNN, ::paddle::platform::CPUPlace,
+                   ops::ConcatMKLDNNOpKernel<float>)

paddle/fluid/operators/concat_op.cc

@@ -13,10 +13,13 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/concat_op.h"
-
 #include <string>
 #include <vector>
 
+#ifdef PADDLE_WITH_MKLDNN
+#include <paddle/fluid/platform/mkldnn_helper.h>
+#endif
+
 namespace paddle {
 namespace operators {
 using framework::Tensor;
@@ -59,6 +62,22 @@ class ConcatOp : public framework::OperatorWithKernel {
     ctx->SetOutputDim("Out", out_dims);
     ctx->ShareLoD("X", /*->*/ "Out");
   }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext &ctx) const override {
+    auto input_data_type =
+        framework::GetDataTypeOfVar(ctx.MultiInputVar("X")[0]);
+
+#ifdef PADDLE_WITH_MKLDNN
+    if (platform::CanMKLDNNBeUsed(ctx)) {
+      return framework::OpKernelType(input_data_type, ctx.GetPlace(),
+                                     framework::DataLayout::kMKLDNN,
+                                     framework::LibraryType::kMKLDNN);
+    }
+#endif
+    return framework::OpKernelType(input_data_type, ctx.GetPlace());
+  }
 };
 
 class ConcatOpMaker : public framework::OpProtoAndCheckerMaker {
@@ -66,6 +85,10 @@ class ConcatOpMaker : public framework::OpProtoAndCheckerMaker {
   void Make() override {
     AddInput("X", "Input tensors of concat operator.").AsDuplicable();
     AddOutput("Out", "Output tensor of concat operator.");
+    AddAttr<bool>(
+        "use_mkldnn",
+        "(bool, default false) Indicates if MKL-DNN kernel will be used")
+        .SetDefault(false);
     AddAttr<int>("axis",
                  "The axis along which the input tensors will be concatenated.")
         .SetDefault(0);

paddle/fluid/operators/cos_sim_op.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/cos_sim_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/crop_op.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/crop_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/distributed/brpc_client.cc

@@ -158,7 +158,7 @@ ChannelQueuePtr BRPCClient::GetChannel(const std::string& ep) {
   for (int i = 0; i < FLAGS_brpc_channel_num; ++i) {
     std::shared_ptr<ChannelContext> c(new ChannelContext());
     if (c->channel.Init(ep.c_str(), &options) != 0) {
-      LOG(ERROR) << "Fail to initialize channel";
+      LOG(FATAL) << "Fail to initialize channel";
       return nullptr;
     }
 

paddle/fluid/operators/distributed/grpc_client.cc

@@ -390,8 +390,7 @@ void GRPCClient::Proceed() {
       VLOG(3) << c->GetVarHandlePtr()->String() << " process";
       c->Process();
     } else if (c->status_.error_code() == grpc::StatusCode::DEADLINE_EXCEEDED) {
-      // FIXME(gongwb): parse error_details?
-      LOG(ERROR) << c->GetVarHandlePtr()->String()
+      LOG(FATAL) << c->GetVarHandlePtr()->String()
                  << " meets grpc error, error_code:" << c->status_.error_code()
                  << " error_message:" << c->status_.error_message()
                  << " error_details:" << c->status_.error_details();

paddle/fluid/operators/dropout_op.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include <thrust/device_ptr.h>
 #include <thrust/iterator/counting_iterator.h>
 #include <thrust/random.h>

paddle/fluid/operators/elementwise/elementwise_add_op.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/elementwise/elementwise_add_op.h"
 #include "paddle/fluid/platform/float16.h"
 

paddle/fluid/operators/elementwise/elementwise_div_op.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/elementwise/elementwise_div_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/elementwise/elementwise_max_op.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/elementwise/elementwise_max_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/elementwise/elementwise_min_op.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/elementwise/elementwise_min_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/elementwise/elementwise_mul_op.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/elementwise/elementwise_mul_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/elementwise/elementwise_pow_op.cu

@@ -8,8 +8,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/elementwise/elementwise_pow_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/elementwise/elementwise_sub_op.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/elementwise/elementwise_sub_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/expand_op.cu

@@ -11,9 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
-
 #include "paddle/fluid/operators/expand_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/gru_unit_op.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/gru_unit_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/hinge_loss_op.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/hinge_loss_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/huber_loss_op.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/huber_loss_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/im2sequence_op.cu

@@ -11,8 +11,6 @@
    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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/im2sequence_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/isfinite_op.cu

@@ -11,8 +11,6 @@
 // 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.
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/isfinite_op.h"
 #include "paddle/fluid/platform/float16.h"
 

paddle/fluid/operators/l1_norm_op.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/l1_norm_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/log_loss_op.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/log_loss_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/math/context_project.cu

@@ -11,9 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
-
 #include "paddle/fluid/operators/math/context_project.h"
 
 namespace paddle {

paddle/fluid/operators/math/jit_kernel_layer_norm.cc

@@ -79,16 +79,16 @@ class LayerNormKernelImpl : public LayerNormKernel<T> {
   }
 };
 
-#define INTRIAVX_FLOAT(isa, block)                                             \
+#define INTRIAVX_FLOAT(isa, jit_block)                                         \
   template <>                                                                  \
-  LayerNormKernelImpl<float, isa, block>::LayerNormKernelImpl(int right)       \
+  LayerNormKernelImpl<float, isa, jit_block>::LayerNormKernelImpl(int right)   \
       : LayerNormKernel<float>() {                                             \
     this->num_ = right;                                                        \
     this->rest_ = this->num_ % YMM_FLOAT_BLOCK;                                \
     this->end_ = this->num_ - this->rest_;                                     \
   }                                                                            \
   template <>                                                                  \
-  void LayerNormKernelImpl<float, platform::avx, block>::Compute(              \
+  void LayerNormKernelImpl<float, isa, jit_block>::Compute(                    \
       float* x, float* out, float* mean, float* var, const float* scale,       \
       const float* bias, int height, const float epsilon) const {              \
     __m256 sum;                                                                \
@@ -97,6 +97,7 @@ class LayerNormKernelImpl : public LayerNormKernel<T> {
     __m256 tmp;                                                                \
     size_t offset;                                                             \
     size_t j;                                                                  \
+    size_t block = YMM_FLOAT_BLOCK;                                            \
     __m256 reverse_num_vec =                                                   \
         _mm256_div_ps(_mm256_set1_ps(1.0), _mm256_set1_ps(this->num_));        \
     __m256 epsilon_vec = _mm256_set1_ps(epsilon);                              \
@@ -221,12 +222,14 @@ INTRIAVX_FLOAT(platform::avx, kEQ8);
 INTRIAVX_FLOAT(platform::avx, kGT8LT16);
 INTRIAVX_FLOAT(platform::avx, kEQ16);
 INTRIAVX_FLOAT(platform::avx, kGT16);
-#endif
-#ifdef __AVX2__
 INTRIAVX_FLOAT(platform::avx2, kEQ8);
 INTRIAVX_FLOAT(platform::avx2, kGT8LT16);
 INTRIAVX_FLOAT(platform::avx2, kEQ16);
 INTRIAVX_FLOAT(platform::avx2, kGT16);
+INTRIAVX_FLOAT(platform::avx512f, kEQ8);
+INTRIAVX_FLOAT(platform::avx512f, kGT8LT16);
+INTRIAVX_FLOAT(platform::avx512f, kEQ16);
+INTRIAVX_FLOAT(platform::avx512f, kGT16);
 #endif
 
 #undef INTRIAVX_FLOAT

paddle/fluid/operators/math/math_function.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include <vector>
 #include "paddle/fluid/framework/data_type.h"
 #include "paddle/fluid/operators/math/blas.h"

paddle/fluid/operators/math/sequence2batch.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/math/sequence2batch.h"
 
 namespace paddle {

paddle/fluid/operators/math/softmax.cu

@@ -11,9 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
-
 #include <vector>
 
 #include "paddle/fluid/operators/math/math_function.h"

paddle/fluid/operators/mean_op.cu

@@ -11,9 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
-
 #include "paddle/fluid/operators/mean_op.h"
 #include "paddle/fluid/platform/float16.h"
 

paddle/fluid/operators/optimizers/adadelta_op.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/optimizers/adadelta_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/optimizers/adagrad_op.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/operators/math/selected_rows_functor.h"
 #include "paddle/fluid/operators/optimizers/adagrad_op.h"

paddle/fluid/operators/optimizers/adam_op.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/optimizers/adam_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/optimizers/adamax_op.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/optimizers/adamax_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/optimizers/decayed_adagrad_op.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/optimizers/decayed_adagrad_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/optimizers/ftrl_op.cu

@@ -10,8 +10,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/optimizers/ftrl_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/optimizers/proximal_adagrad_op.cu

@@ -10,8 +10,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/optimizers/proximal_adagrad_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/optimizers/proximal_gd_op.cu

@@ -10,8 +10,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/optimizers/proximal_gd_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/optimizers/rmsprop_op.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/optimizers/rmsprop_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/pad_constant_like_op.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/pad_constant_like_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/pad_op.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/pad_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/sequence_ops/sequence_pool_op.cu

@@ -11,9 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
-
 #include "paddle/fluid/operators/sequence_ops/sequence_pool_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/smooth_l1_loss_op.cu

@@ -11,9 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
-
 #include "paddle/fluid/operators/smooth_l1_loss_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/softmax_with_cross_entropy_op.cu

@@ -11,9 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
-
 #include <cub/cub.cuh>
 #include "paddle/fluid/operators/math/cross_entropy.h"
 #include "paddle/fluid/operators/softmax_with_cross_entropy_op.h"

paddle/fluid/operators/split_selected_rows_op.h

@@ -72,10 +72,11 @@ class SplitSelectedRowsOpKernel : public framework::OpKernel<T> {
     for (size_t i = 0; i < outs_rows_idx.size(); ++i) {
       auto rows_idx = outs_rows_idx[i];
       outs[i]->set_height(height_sections[i]);
+      auto dims = x->GetCompleteDims();
+      dims[0] = rows_idx.size();
+      outs[i]->mutable_value()->mutable_data<T>(dims, x->place());
+      outs[i]->mutable_rows()->clear();
       if (rows_idx.size() > 0) {
-        auto dims = x->GetCompleteDims();
-        dims[0] = rows_idx.size();
-        outs[i]->mutable_value()->mutable_data<T>(dims, x->place());
         for (auto idx : rows_idx) {
           outs[i]->mutable_rows()->push_back(idx - abs_sections[i]);
         }
@@ -98,6 +99,8 @@ class SplitSelectedRowsOpKernel : public framework::OpKernel<T> {
           }
         }
       }
+      PADDLE_ENFORCE_EQ(rows_idx.size(), outs[i]->rows().size(),
+                        "rows should has the same size with tensor dim 0");
     }
   }
 };

paddle/fluid/operators/squared_l2_distance_op.cu

@@ -11,9 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
-
 #include "paddle/fluid/operators/squared_l2_distance_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/squared_l2_norm_op.cu

@@ -11,8 +11,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/squared_l2_norm_op.h"
 
 namespace ops = paddle::operators;

paddle/fluid/operators/sum_op.cu

@@ -8,8 +8,6 @@ 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. */
-
-#define EIGEN_USE_GPU
 #include "paddle/fluid/operators/sum_op.h"
 #include "paddle/fluid/platform/float16.h"
 

paddle/fluid/platform/cuda_helper_test.cu

@@ -93,7 +93,7 @@ TEST(CudaAtomic, float16) {
 
 // unalignment of uint8
 void TestUnalign(size_t num, const int shift_bit) {
-  PADDLE_ENFORCE(num % 2 == 0, "must be a multiple of 2");
+  ASSERT_EQ(num % 2, 0);
   float16 *in1, *in2, *out;
   float16 *d_in1, *d_in2;
   size_t size = sizeof(uint8_t) * (num + shift_bit);

paddle/fluid/platform/device_context.cc

@@ -220,6 +220,40 @@ CUDADeviceContext::CUDADeviceContext(CUDAPlace place)
   LOG_FIRST_N(WARNING, 1) << "device: " << place_.device
                           << ", cuDNN Version: " << cudnn_dso_ver / 1000 << "."
                           << (cudnn_dso_ver % 100) / 10 << ".";
+
+  {
+    // Check CUDA/CUDNN version compatiblity
+    auto local_cuda_version = runtime_version_ / 100;
+    auto compile_cuda_version = CUDA_VERSION / 100;
+    if (local_cuda_version < compile_cuda_version) {
+      LOG_FIRST_N(WARNING, 1)
+          << "WARNING: device: " << place_.device
+          << ". The installed Paddle is compiled with CUDA "
+          << compile_cuda_version / 10 << "." << compile_cuda_version % 10
+          << ", but CUDA runtime version in your machine is "
+          << local_cuda_version / 10 << "." << local_cuda_version % 10
+          << ", which may cause serious incompatible bug. "
+          << "Please recompile or reinstall Paddle with compatible CUDA "
+             "version.";
+    }
+
+    if (dynload::HasCUDNN()) {
+      auto local_cudnn_version = cudnn_dso_ver / 100;
+      auto compile_cudnn_version = CUDNN_VERSION / 100;
+      if (local_cuda_version < compile_cuda_version) {
+        LOG_FIRST_N(WARNING, 1)
+            << "WARNING: device: " << place_.device
+            << ". The installed Paddle is compiled with CUDNN "
+            << compile_cudnn_version / 10 << "." << compile_cudnn_version % 10
+            << ", but CUDNN version in your machine is "
+            << local_cudnn_version / 10 << "." << local_cudnn_version % 10
+            << ", which may cause serious incompatible bug. "
+            << "Please recompile or reinstall Paddle with compatible CUDNN "
+               "version.";
+      }
+    }
+  }
+
   callback_manager_.reset(new StreamCallbackManager(stream_));
 }
 

paddle/fluid/platform/device_context.h

@@ -21,7 +21,6 @@ limitations under the License. */
 #include "paddle/fluid/platform/dynload/cublas.h"
 #include "paddle/fluid/platform/dynload/cudnn.h"
 #include "paddle/fluid/platform/gpu_info.h"
-#define EIGEN_USE_GPU
 #endif
 
 #ifdef PADDLE_WITH_MKLDNN

paddle/fluid/platform/enforce.h

@@ -62,45 +62,54 @@ inline std::string demangle(std::string name) { return name; }
 #endif
 
 struct EnforceNotMet : public std::exception {
-  std::exception_ptr exp_;
   std::string err_str_;
-  EnforceNotMet(std::exception_ptr e, const char* f, int l) : exp_(e) {
-    static constexpr int TRACE_STACK_LIMIT = 100;
+  EnforceNotMet(std::exception_ptr e, const char* f, int l) {
     try {
-      std::rethrow_exception(exp_);
-    } catch (const std::exception& exp) {
-      std::ostringstream sout;
+      std::rethrow_exception(e);
+    } catch (std::exception& e) {
+      Init(e.what(), f, l);
+    }
+  }
 
-      sout << string::Sprintf("%s at [%s:%d]", exp.what(), f, l) << std::endl;
-      sout << "PaddlePaddle Call Stacks: " << std::endl;
+  template <typename... ARGS>
+  EnforceNotMet(const char* f, int l, ARGS... args) {
+    Init(string::Sprintf(args...), f, l);
+  }
+
+  const char* what() const noexcept override { return err_str_.c_str(); }
+
+ private:
+  template <typename StrType>
+  inline void Init(StrType what, const char* f, int l) {
+    static constexpr int TRACE_STACK_LIMIT = 100;
+    std::ostringstream sout;
+
+    sout << string::Sprintf("%s at [%s:%d]", what, f, l) << std::endl;
+    sout << "PaddlePaddle Call Stacks: " << std::endl;
 #if !defined(_WIN32)
-      void* call_stack[TRACE_STACK_LIMIT];
-      auto size = backtrace(call_stack, TRACE_STACK_LIMIT);
-      auto symbols = backtrace_symbols(call_stack, size);
-
-      Dl_info info;
-      for (int i = 0; i < size; ++i) {
-        if (dladdr(call_stack[i], &info) && info.dli_sname) {
-          auto demangled = demangle(info.dli_sname);
-          auto addr_offset = static_cast<char*>(call_stack[i]) -
-                             static_cast<char*>(info.dli_saddr);
-          sout << string::Sprintf("%-3d %*0p %s + %zd\n", i,
-                                  2 + sizeof(void*) * 2, call_stack[i],
-                                  demangled, addr_offset);
-        } else {
-          sout << string::Sprintf("%-3d %*0p\n", i, 2 + sizeof(void*) * 2,
-                                  call_stack[i]);
-        }
+    void* call_stack[TRACE_STACK_LIMIT];
+    auto size = backtrace(call_stack, TRACE_STACK_LIMIT);
+    auto symbols = backtrace_symbols(call_stack, size);
+    Dl_info info;
+    for (int i = 0; i < size; ++i) {
+      if (dladdr(call_stack[i], &info) && info.dli_sname) {
+        auto demangled = demangle(info.dli_sname);
+        auto addr_offset = static_cast<char*>(call_stack[i]) -
+                           static_cast<char*>(info.dli_saddr);
+        sout << string::Sprintf("%-3d %*0p %s + %zd\n", i,
+                                2 + sizeof(void*) * 2, call_stack[i], demangled,
+                                addr_offset);
+      } else {
+        sout << string::Sprintf("%-3d %*0p\n", i, 2 + sizeof(void*) * 2,
+                                call_stack[i]);
       }
-      free(symbols);
+    }
+    free(symbols);
 #else
-      sout << "Windows not support stack backtrace yet.";
+    sout << "Windows not support stack backtrace yet.";
 #endif
-      err_str_ = sout.str();
-    }
+    err_str_ = sout.str();
   }
-
-  const char* what() const noexcept { return err_str_.c_str(); }
 };
 
 struct EOFException : public std::exception {
@@ -242,13 +251,8 @@ inline void throw_on_error(T e) {
   throw_on_error(e, "");
 }
 
-#define PADDLE_THROW(...)                                              \
-  do {                                                                 \
-    throw ::paddle::platform::EnforceNotMet(                           \
-        std::make_exception_ptr(                                       \
-            std::runtime_error(paddle::string::Sprintf(__VA_ARGS__))), \
-        __FILE__, __LINE__);                                           \
-  } while (false)
+#define PADDLE_THROW(...) \
+  throw ::paddle::platform::EnforceNotMet(__FILE__, __LINE__, __VA_ARGS__)
 
 #ifndef REPLACE_ENFORCE_GLOG
 #define PADDLE_ENFORCE(...)                                             \

paddle/fluid/platform/float16.h

@@ -71,9 +71,6 @@ struct float16;
 }  // namespace platform
 }  // namespace paddle
 
-// NOTE():
-// Do not move the eigen.h header, otherwise the eigen_vector<bool> will failed.
-#include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/platform/hostdevice.h"
 #include "unsupported/Eigen/CXX11/Tensor"
 

paddle/fluid/pybind/CMakeLists.txt

 
-set(PYBIND_DEPS pybind python proto_desc memory executor async_executor prune feed_fetch_method pass_builder parallel_executor profiler)
-set(PYBIND_SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc async_executor_py.cc)
+set(PYBIND_DEPS pybind python proto_desc memory executor async_executor prune feed_fetch_method pass_builder parallel_executor profiler layer)
+set(PYBIND_SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc async_executor_py.cc imperative.cc)
+
 if(WITH_PYTHON)
   if(WITH_AMD_GPU)
     hip_library(paddle_pybind SHARED

paddle/fluid/pybind/imperative.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/pybind/imperative.h"
+#include "paddle/fluid/framework/block_desc.h"
+#include "paddle/fluid/framework/scope.h"
+#include "paddle/fluid/imperative/tracer.h"
+
+namespace paddle {
+namespace pybind {
+
+// Bind Methods
+void BindTracer(pybind11::module *m) {
+  pybind11::class_<imperative::Tracer>(*m, "Tracer", "")
+      .def("__init__",
+           [](imperative::Tracer &self, framework::BlockDesc *root_block) {
+             new (&self) imperative::Tracer(root_block);
+           })
+      .def("trace", &imperative::Tracer::Trace)
+      .def("get_scope", &imperative::Tracer::GetScope,
+           pybind11::return_value_policy::reference);
+}
+
+}  // namespace pybind
+}  // namespace paddle

paddle/fluid/pybind/imperative.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. */
+#pragma once
+
+#include <Python.h>
+#include <vector>
+#include "paddle/fluid/imperative/layer.h"
+#include "pybind11/pybind11.h"
+#include "pybind11/stl.h"
+
+namespace paddle {
+namespace pybind {
+
+class PyLayer : public imperative::Layer {
+ public:
+  using imperative::Layer::Layer;  // Inherit constructors
+
+  std::vector<imperative::VarBase> Forward(
+      const std::vector<imperative::VarBase>& inputs) override {
+    PYBIND11_OVERLOAD(std::vector<imperative::VarBase>, Layer, Forward,
+                      inputs);  // NOLINT
+  }
+
+  void Backward() override {
+    PYBIND11_OVERLOAD(void, Layer, Backward, );  // NOLINT
+  }
+};
+
+class PyOpBase : public imperative::OpBase {
+ public:
+  using imperative::OpBase::OpBase;  // Inherit constructors
+};
+
+class PyVarBase : public imperative::VarBase {
+ public:
+  using imperative::VarBase::VarBase;  // Inherit constructors
+};
+
+void BindTracer(pybind11::module* m);
+
+}  // namespace pybind
+}  // namespace paddle

paddle/fluid/pybind/pybind.cc

@@ -34,6 +34,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/reader.h"
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/framework/version.h"
+#include "paddle/fluid/imperative/layer.h"
 #include "paddle/fluid/memory/allocation/allocator_strategy.h"
 #include "paddle/fluid/operators/activation_op.h"
 #include "paddle/fluid/operators/py_func_op.h"
@@ -46,6 +47,7 @@ limitations under the License. */
 #include "paddle/fluid/pybind/async_executor_py.h"
 #include "paddle/fluid/pybind/const_value.h"
 #include "paddle/fluid/pybind/exception.h"
+#include "paddle/fluid/pybind/imperative.h"
 #include "paddle/fluid/pybind/protobuf.h"
 #include "paddle/fluid/pybind/pybind.h"  // NOLINT
 #include "paddle/fluid/pybind/recordio.h"
@@ -107,6 +109,42 @@ PYBIND11_MODULE(core, m) {
         return paddle::operators::AppendPythonCallableObjectAndReturnId(py_obj);
       });
 
+  py::class_<imperative::VarBase, PyVarBase>(m, "VarBase", R"DOC()DOC")
+      .def(py::init<>())
+      .def("_run_backward",
+           [](imperative::VarBase &self, framework::Scope *scope) {
+             self.RunBackward(scope);
+           })
+      .def("_grad", &imperative::VarBase::Grad)
+      .def_property(
+          "desc",
+          [](const imperative::VarBase &self) { return self.var_desc_; },
+          [](imperative::VarBase &self, framework::VarDesc *var_desc) {
+            self.var_desc_ = var_desc;
+          },
+          py::return_value_policy::reference);
+
+  py::class_<imperative::OpBase, PyOpBase>(m, "OpBase", R"DOC()DOC")
+      .def(py::init<>())
+      .def_property(
+          "desc", [](const imperative::OpBase &self) { return self.op_desc_; },
+          [](imperative::OpBase &self, framework::OpDesc *op_desc) {
+            if (op_desc) {
+              self.op_desc_ = op_desc;
+            }
+          },
+          py::return_value_policy::reference);
+
+  py::class_<imperative::Layer, PyLayer /* <--- trampoline*/> layer(m, "Layer");
+  layer.def(py::init<>())
+      .def("forward",
+           [](imperative::Layer &self,
+              const std::vector<imperative::VarBase> &inputs) {
+             return self.Forward(inputs);
+           })
+      .def("backward", &imperative::Layer::Backward);
+  BindTracer(&m);
+
   py::class_<Tensor>(m, "Tensor", py::buffer_protocol())
       .def_buffer(
           [](Tensor &self) -> py::buffer_info { return CastToPyBuffer(self); })
@@ -305,6 +343,8 @@ PYBIND11_MODULE(core, m) {
       .def("get_tensor",
            [](SelectedRows &self) { return self.mutable_value(); },
            py::return_value_policy::reference)
+      .def("numel",
+           [](SelectedRows &self) -> int64_t { return self.value().numel(); })
       .def("set_height", &SelectedRows::set_height)
       .def("height", &SelectedRows::height)
       .def("set_rows",
@@ -622,6 +662,7 @@ All parameter, weight, gradient are variables in Paddle.
 
   m.def("set_feed_variable", framework::SetFeedVariable);
   m.def("get_fetch_variable", framework::GetFetchVariable);
+  m.def("get_variable_tensor", framework::GetVariableTensor);
 
   m.def("_is_program_version_supported", IsProgramVersionSupported);
 

paddle/fluid/pybind/tensor_py.h

@@ -182,7 +182,7 @@ inline void PyCPUTensorSetFromArray(
     paddle::platform::CPUPlace place) {
   std::vector<int64_t> dims;
   dims.reserve(array.ndim());
-  for (size_t i = 0; i < array.ndim(); ++i) {
+  for (int i = 0; i < array.ndim(); ++i) {
     dims.push_back(static_cast<int>(array.shape()[i]));
   }
 

python/paddle/dataset/image.py

@@ -32,11 +32,28 @@ the image layout as follows.
 
 from __future__ import print_function
 
+import six
 import numpy as np
-try:
-    import cv2
-except ImportError:
-    cv2 = None
+# FIXME(minqiyang): this is an ugly fix for the numpy bug reported here
+# https://github.com/numpy/numpy/issues/12497
+if six.PY3:
+    import subprocess
+    import sys
+    import_cv2_proc = subprocess.Popen(
+        [sys.executable, "-c", "import cv2"],
+        stdout=subprocess.PIPE,
+        stderr=subprocess.PIPE)
+    out, err = import_cv2_proc.communicate()
+    retcode = import_cv2_proc.poll()
+    if retcode != 0:
+        cv2 = None
+    else:
+        import cv2
+else:
+    try:
+        import cv2
+    except ImportError:
+        cv2 = None
 import os
 import tarfile
 import six.moves.cPickle as pickle

python/paddle/fluid/__init__.py

@@ -34,6 +34,7 @@ from . import io
 from . import evaluator
 from . import initializer
 from . import layers
+from . import imperative
 from . import contrib
 from . import nets
 from . import optimizer
@@ -67,6 +68,7 @@ __all__ = framework.__all__ + executor.__all__ + \
         'initializer',
         'layers',
         'contrib',
+        'imperative',
         'transpiler',
         'nets',
         'optimizer',
@@ -125,7 +127,8 @@ def __bootstrap__():
         'use_ngraph', 'initial_cpu_memory_in_mb', 'init_allocated_mem',
         'free_idle_memory', 'paddle_num_threads', "dist_threadpool_size",
         'eager_delete_tensor_gb', 'allocator_strategy',
-        'reader_queue_speed_test_mode', 'print_sub_graph_dir'
+        'reader_queue_speed_test_mode', 'print_sub_graph_dir',
+        'pe_profile_fname'
     ]
     if 'Darwin' not in sysstr:
         read_env_flags.append('use_pinned_memory')

python/paddle/fluid/average.py

@@ -48,6 +48,7 @@ class WeightedAverage(object):
 
     Examples:
         .. code-block:: python
+
             avg = fluid.average.WeightedAverage()
             avg.add(value=2.0, weight=1)
             avg.add(value=4.0, weight=2)

python/paddle/fluid/framework.py

@@ -18,6 +18,7 @@ import collections
 import contextlib
 import re
 import six
+import sys
 
 import numpy as np
 
@@ -49,6 +50,16 @@ GRAD_VAR_SUFFIX = core.kGradVarSuffix()
 ZERO_VAR_SUFFIX = core.kZeroVarSuffix()
 CONTROL_DEP_VAR_PREFIX = core.kControlDepVarName()
 
+_imperative_tracer_ = None
+
+
+def _in_imperative_mode():
+    return _imperative_tracer_ is not None
+
+
+def _imperative_tracer():
+    return _imperative_tracer_
+
 
 class NameScope(object):
     def __init__(self, name="", parent=None):
@@ -345,6 +356,21 @@ class Variable(object):
         self.op = None
         self.stop_gradient = stop_gradient
         self.is_data = is_data
+        if _in_imperative_mode():
+            self._ivar = core.VarBase()
+            self._ivar.desc = self.desc
+
+    def _numpy(self):
+        scope = _imperative_tracer().get_scope(self.block.desc)
+        tensor = core.get_variable_tensor(scope, self.desc.name())
+        return np.array(tensor)
+
+    def _backward(self):
+        scope = _imperative_tracer().get_scope(self.block.desc)
+        self._ivar._run_backward(scope)
+
+    def _gradient(self):
+        return np.array(self._ivar._grad())
 
     def __str__(self):
         return self.to_string(True)
@@ -655,6 +681,23 @@ class Operator(object):
         if self._has_kernel(type):
             self.desc.infer_var_type(self.block.desc)
             self.desc.infer_shape(self.block.desc)
+        if _in_imperative_mode():
+            self.iop = core.OpBase()
+            self.iop.desc = self.desc
+            self.inputs = []
+            if inputs is not None:
+                for inp in inputs.values():
+                    if isinstance(inp, Variable):
+                        self.inputs.append(inp)
+                    elif isinstance(inp, list) or isinstance(inp, tuple):
+                        self.inputs.extend(inp[:])
+            self.outputs = []
+            if outputs is not None:
+                for out in outputs.values():
+                    if isinstance(out, Variable):
+                        self.outputs.append(out)
+                    elif isinstance(out, list) or isinstance(out, tuple):
+                        self.outputs.extend(out[:])
 
     def _has_kernel(self, op_type):
         return op_type not in self.OP_WITHOUT_KERNEL_SET
@@ -1041,19 +1084,15 @@ class Block(object):
             raise ValueError("var %s not in this block" % name)
         return v
 
-    def _var_recursive(self, name):
+    def _find_var_recursive(self, name):
         """
         Get a Variable by name from this block recursively.
 
         Args:
             name(str): the Variable's name.
 
-        Raises:
-            ValueError: this block and this parent block doesn't
-                have a Variable with the giving name.
-
         Returns:
-            Variable: the Variable with the giving name.
+            Variable: the Variable with the giving name. Or None if not found.
         """
         frontier = list()
         visited = set()
@@ -1079,8 +1118,27 @@ class Block(object):
                 frontier.append(prog.block(cur.forward_block_idx))
 
             visited.add(id(cur))
+        return None
 
-        raise ValueError("Var {0} is not found recursively".format(name))
+    def _var_recursive(self, name):
+        """
+        Get a Variable by name from this block recursively.
+
+        Args:
+            name(str): the Variable's name.
+
+        Raises:
+            ValueError: this block and this parent block doesn't
+                have a Variable with the giving name.
+
+        Returns:
+            Variable: the Variable with the giving name.
+        """
+        var = self._find_var_recursive(name)
+        if var:
+            return var
+        else:
+            raise ValueError("Var {0} is not found recursively".format(name))
 
     def all_parameters(self):
         return list(self.iter_parameters())
@@ -1206,6 +1264,9 @@ class Block(object):
         """
         op_desc = self.desc.append_op()
         op = Operator(block=self, desc=op_desc, *args, **kwargs)
+        if _in_imperative_mode():
+            _imperative_tracer().trace(op.iop, [v._ivar for v in op.inputs],
+                                       [v._ivar for v in op.outputs], self.desc)
         self.ops.append(op)
         return op
 
@@ -2210,3 +2271,12 @@ def _get_var(name, program=None):
     assert isinstance(program, Program)
 
     return program.global_block().var(name)
+
+
+@contextlib.contextmanager
+def _imperative_guard(tracer):
+    global _imperative_tracer_
+    tmp_trace = _imperative_tracer_
+    _imperative_tracer_ = tracer
+    yield
+    _imperative_tracer_ = tmp_trace

python/paddle/fluid/imperative/__init__.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.
+
+from __future__ import print_function
+
+from . import base
+from .base import *
+
+from . import layers
+from .layers import *
+
+__all__ = []
+__all__ += layers.__all__
+__all__ += base.__all__

python/paddle/fluid/imperative/base.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 contextlib
+import numpy as np
+
+from paddle.fluid import core
+from paddle.fluid import framework
+
+__all__ = ['enabled', 'guard', 'to_variable']
+
+
+def enabled():
+    return framework._in_imperative_mode()
+
+
+@contextlib.contextmanager
+def guard():
+    train = framework.Program()
+    startup = framework.Program()
+    tracer = core.Tracer(train.current_block().desc)
+    with framework.program_guard(train, startup):
+        with framework.unique_name.guard():
+            with framework._imperative_guard(tracer):
+                yield
+
+
+def to_variable(value, block=None):
+    if isinstance(value, np.ndarray):
+        if not block:
+            block = framework.default_main_program().current_block()
+        py_var = framework.Variable(
+            block,
+            type=core.VarDesc.VarType.LOD_TENSOR,
+            name=None,
+            shape=value.shape,
+            dtype=value.dtype)
+        scope = framework._imperative_tracer().get_scope(block.desc)
+        var = scope.var(py_var.name)
+        tensor = var.get_tensor()
+        tensor.set(value, core.CPUPlace())
+        return py_var
+    elif isinstance(value, framework.Variable):
+        return value
+    else:
+        raise ValueError("Unsupported type %s" % type(value))

python/paddle/fluid/imperative/layers.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 contextlib
+import sys
+import numpy as np
+
+from paddle.fluid import core
+from paddle.fluid import framework
+from paddle.fluid.imperative import base
+
+__all__ = ['PyLayer']
+
+
+class PyLayer(core.Layer):
+    def __init__(self):
+        pass
+
+    def __call__(self, inputs):
+        # TODO(panyx0718): Support declarative mode as well.
+        assert base.enabled()
+        if not isinstance(inputs, list) and not isinstance(inputs, tuple):
+            inputs = [inputs]
+
+        var_inputs = []
+        for x in inputs:
+            py_var = base.to_variable(x)
+            var_inputs.append(py_var)
+        outputs = self.forward(var_inputs)
+        return outputs
+
+    def forward(self, inputs):
+        return []

python/paddle/fluid/layer_helper.py

@@ -17,10 +17,13 @@ from __future__ import print_function
 import copy
 import itertools
 import six
+import sys
+import numpy as np
 
 from .framework import Variable, Parameter, default_main_program, default_startup_program, dtype_is_floating
 from . import unique_name
 from paddle.fluid.initializer import Constant, Xavier
+from paddle.fluid.imperative import base
 from .param_attr import ParamAttr, WeightNormParamAttr
 from . import core
 from six.moves import zip
@@ -46,23 +49,21 @@ class LayerHelper(object):
     def startup_program(self):
         return default_startup_program()
 
+    def to_variable(self, x):
+        return base.to_variable(x, self.main_program.current_block())
+
     def append_op(self, *args, **kwargs):
         return self.main_program.current_block().append_op(*args, **kwargs)
 
     def multiple_input(self, input_param_name='input'):
         inputs = self.kwargs.get(input_param_name, [])
-        type_error = TypeError(
-            "Input of {0} layer should be Variable or sequence of Variable".
-            format(self.layer_type))
-        if isinstance(inputs, Variable):
-            inputs = [inputs]
-        elif not isinstance(inputs, list) and not isinstance(inputs, tuple):
-            raise type_error
+        ret = []
+        if isinstance(inputs, list) or isinstance(inputs, tuple):
+            for inp in inputs:
+                ret.append(self.to_variable(inp))
         else:
-            for each in inputs:
-                if not isinstance(each, Variable):
-                    raise type_error
-        return inputs
+            ret.append(self.to_variable(inputs))
+        return ret
 
     def input(self, input_param_name='input'):
         inputs = self.multiple_input(input_param_name)

python/paddle/fluid/layers/control_flow.py

@@ -717,8 +717,9 @@ class While(object):
 
         out_vars = []
         for inner_out_name in inner_outputs:
-            if inner_out_name in parent_block.vars:
-                out_vars.append(parent_block.var(inner_out_name))
+            inner_var = parent_block._find_var_recursive(inner_out_name)
+            if inner_var:
+                out_vars.append(inner_var)
 
         step_scope = parent_block.create_var(
             type=core.VarDesc.VarType.STEP_SCOPES)
@@ -1264,10 +1265,11 @@ class ConditionalBlock(object):
             if each_name not in input_set
         ]
 
-        out_list = [
-            parent_block.var(var_name) for var_name in parent_block.vars
-            if var_name in intermediate
-        ]
+        out_list = []
+        for inner_out_name in intermediate:
+            inner_var = parent_block._find_var_recursive(inner_out_name)
+            if inner_var:
+                out_list.append(inner_var)
 
         step_scope = parent_block.create_var(
             type=core.VarDesc.VarType.STEP_SCOPES)

python/paddle/fluid/layers/nn.py

@@ -43,6 +43,7 @@ __all__ = [
     'crf_decoding',
     'cos_sim',
     'cross_entropy',
+    'bpr_loss',
     'square_error_cost',
     'chunk_eval',
     'sequence_conv',
@@ -1351,6 +1352,44 @@ def cross_entropy(input, label, soft_label=False, ignore_index=kIgnoreIndex):
     return out
 
 
+def bpr_loss(input, label, name=None):
+    """
+    Bayesian Personalized Ranking Loss Operator.
+
+    This operator belongs to pairwise ranking loss. Label is the desired item.
+    The loss at a given point in one session is defined as:
+    $Y[i] = -\frac{1}{N_{i}-1} * \sum_{0\le j<N_{i},~ j\neq Label[i]}\log(\sigma(X[i, Label[i]]-X[i, j]))$
+
+    Learn more details by reading paper <session-based recommendations with recurrent
+    neural networks>(https://arxiv.org/abs/1511.06939)
+
+    Args:
+        input (Variable|list):  a 2-D tensor with shape [N x D], where N is the
+                                batch size and D is the number of classes.
+                                This input is not probability but logits.
+        label (Variable|list):  the ground truth which is a 2-D tensor.  `label`
+                                is a tensor<int64> with shape [N x 1].
+        name (str|None):        A name for this layer(optional). If set None, the
+                                layer will be named automatically. Default: None.
+    Returns:
+        A 2-D tensor with shape [N x 1], the bpr loss.
+
+    Examples:
+        .. code-block:: python
+
+          cost = fluid.layers.bpr_loss(input=predict, label=label)
+    """
+
+    helper = LayerHelper('bpr_loss', **locals())
+    out = helper.create_variable_for_type_inference(dtype=input.dtype)
+    helper.append_op(
+        type='bpr_loss',
+        inputs={'X': [input],
+                'Label': [label]},
+        outputs={'Y': [out]})
+    return out
+
+
 def square_error_cost(input, label):
     """
     **Square error cost layer**
@@ -6626,7 +6665,8 @@ def relu(x, name=None):
     helper = LayerHelper('relu', **locals())
     dtype = helper.input_dtype(input_param_name='x')
     out = helper.create_variable_for_type_inference(dtype)
-    helper.append_op(type="relu", inputs={"X": x}, outputs={"Out": out})
+    helper.append_op(
+        type="relu", inputs={"X": helper.input('x')}, outputs={"Out": out})
     return out
 
 

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

@@ -93,7 +93,7 @@ class TestDistMnist2x2(TestDistRunnerBase):
         # TODO(typhoonzero): fix distributed adam optimizer
         # opt = fluid.optimizer.AdamOptimizer(
         #     learning_rate=0.001, beta1=0.9, beta2=0.999)
-        opt = fluid.optimizer.Momentum(learning_rate=0.001, momentum=0.9)
+        opt = fluid.optimizer.Momentum(learning_rate=self.lr, momentum=0.9)
 
         # Reader
         train_reader = paddle.batch(

python/paddle/fluid/tests/unittests/test_bpr_loss_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.
+
+from __future__ import print_function
+
+import unittest
+import numpy as np
+from op_test import OpTest, randomize_probability
+
+
+class TestBprLossOp1(OpTest):
+    """Test BprLoss with discrete one-hot labels.
+    """
+
+    def setUp(self):
+        self.op_type = "bpr_loss"
+        batch_size = 40
+        class_num = 5
+        X = randomize_probability(batch_size, class_num, dtype='float64')
+        label = np.random.randint(0, class_num, (batch_size, 1), dtype="int64")
+        bpr_loss_result = []
+        for i in range(batch_size):
+            sum = 0.0
+            for j in range(class_num):
+                if j == label[i][0]:
+                    continue
+                sum += (-np.log(1.0 + np.exp(X[i][j] - X[i][label[i][0]])))
+            bpr_loss_result.append(-sum / (class_num - 1))
+        bpr_loss = np.asmatrix([[x] for x in bpr_loss_result], dtype="float64")
+        self.inputs = {"X": X, "Label": label}
+        self.outputs = {"Y": bpr_loss}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(["X"], "Y", numeric_grad_delta=0.001)
+
+
+if __name__ == "__main__":
+    unittest.main()

python/paddle/fluid/tests/unittests/test_concat_mkldnn_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.
+
+from __future__ import print_function
+
+import unittest
+from test_concat_op import TestConcatOp, TestConcatOp2, TestConcatOp3
+
+
+class TestMKLDNNConcatOp(TestConcatOp):
+    def setUp(self):
+        super(TestMKLDNNConcatOp, self).setUp()
+        self.attrs["use_mkldnn"] = True
+        self._cpu_only = True
+
+    def test_check_grad(self):
+        pass
+
+    def init_kernel_type(self):
+        self.use_mkldnn = True
+
+
+class TestMKLDNNConcatOp2(TestConcatOp2):
+    def setUp(self):
+        super(TestMKLDNNConcatOp2, self).setUp()
+        self.attrs["use_mkldnn"] = True
+        self._cpu_only = True
+
+    def test_check_grad(self):
+        pass
+
+    def init_kernel_type(self):
+        self.use_mkldnn = True
+
+
+class TestMKLDNNConcatOp3(TestConcatOp3):
+    def setUp(self):
+        super(TestMKLDNNConcatOp3, self).setUp()
+        self.attrs["use_mkldnn"] = True
+        self._cpu_only = True
+
+    def test_check_grad(self):
+        pass
+
+    def init_kernel_type(self):
+        self.use_mkldnn = True
+
+
+if __name__ == '__main__':
+    unittest.main()

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

@@ -32,7 +32,7 @@ DEFAULT_BATCH_SIZE = 2
 
 
 class TestDistRunnerBase(object):
-    def get_model(self, batch_size=DEFAULT_BATCH_SIZE):
+    def get_model(self, batch_size=DEFAULT_BATCH_SIZE, lr=0.1):
         raise NotImplementedError(
             "get_model should be implemented by child classes.")
 
@@ -56,6 +56,7 @@ class TestDistRunnerBase(object):
         return t
 
     def run_pserver(self, args):
+        self.lr = args.lr
         self.get_model(batch_size=args.batch_size)
         # NOTE: pserver should not call memory optimize
         t = self.get_transpiler(args.trainer_id,
@@ -71,6 +72,7 @@ class TestDistRunnerBase(object):
         exe.run(pserver_prog)
 
     def run_trainer(self, args):
+        self.lr = args.lr
         test_program, avg_cost, train_reader, test_reader, batch_acc, predict = \
             self.get_model(batch_size=args.batch_size)
 
@@ -189,6 +191,7 @@ def runtime_main(test_class):
     parser.add_argument(
         '--use_reader_alloc', action='store_true', required=False)
     parser.add_argument('--batch_size', required=False, type=int, default=2)
+    parser.add_argument('--lr', required=False, type=float, default=0.001)
     parser.add_argument(
         '--batch_merge_repeat', required=False, type=int, default=1)
 
@@ -234,6 +237,7 @@ class TestDistBase(unittest.TestCase):
         self._dc_asgd = False  # must use with async mode
         self._use_reader_alloc = True
         self._nccl2_mode = False
+        self._lr = 0.001
         self._setup_config()
         self._after_setup_config()
 
@@ -284,7 +288,8 @@ class TestDistBase(unittest.TestCase):
                    batch_size=DEFAULT_BATCH_SIZE,
                    batch_merge_repeat=1):
 
-        cmd = "%s %s --role trainer" % (self._python_interp, model)
+        cmd = "%s %s --role trainer --lr %f" % (self._python_interp, model,
+                                                self._lr)
         if batch_size != DEFAULT_BATCH_SIZE:
             cmd += " --batch_size %d" % batch_size
         if batch_merge_repeat > 1:
@@ -330,13 +335,13 @@ class TestDistBase(unittest.TestCase):
 
         ps0_ep, ps1_ep = self._ps_endpoints.split(",")
 
-        tr_cmd = "%s %s --role trainer --endpoints %s --trainer_id %d --current_endpoint %s --trainers %d --update_method pserver"
+        tr_cmd = "%s %s --role trainer --endpoints %s --trainer_id %d --current_endpoint %s --trainers %d --update_method pserver --lr %f"
         tr0_cmd = tr_cmd % \
                   (self._python_interp, model, self._ps_endpoints,
-                   0, ps0_ep, self._trainers)
+                   0, ps0_ep, self._trainers, self._lr)
         tr1_cmd = tr_cmd % \
                   (self._python_interp, model, self._ps_endpoints,
-                   1, ps1_ep, self._trainers)
+                   1, ps1_ep, self._trainers, self._lr)
 
         if self._sync_mode:
             tr0_cmd += " --sync_mode"
@@ -378,6 +383,18 @@ class TestDistBase(unittest.TestCase):
             stderr=tr1_pipe,
             env=env1)
 
+        # Wait until trainer process terminate
+        while True:
+            stat0 = tr0_proc.poll()
+            time.sleep(0.1)
+            if stat0 is not None:
+                break
+        while True:
+            stat1 = tr1_proc.poll()
+            time.sleep(0.1)
+            if stat1 is not None:
+                break
+
         tr0_out, tr0_err = tr0_proc.communicate()
         tr1_out, tr1_err = tr1_proc.communicate()
 
@@ -390,11 +407,21 @@ class TestDistBase(unittest.TestCase):
         ps0.terminate()
         ps1.terminate()
 
+        # print server log
+        with open("/tmp/ps0_err.log", "r") as fn:
+            sys.stderr.write("ps0 stderr: %s\n" % fn.read())
+        with open("/tmp/ps1_err.log", "r") as fn:
+            sys.stderr.write("ps1 stderr: %s\n" % fn.read())
+
         # print log
-        sys.stderr.write('trainer 0 stdout: %s\n' % pickle.loads(tr0_out))
-        sys.stderr.write('trainer 0 stderr: %s\n' % tr0_err)
-        sys.stderr.write('trainer 1 stdout: %s\n' % pickle.loads(tr1_out))
-        sys.stderr.write('trainer 1 stderr: %s\n' % tr1_err)
+        if stat0 == 0:
+            sys.stderr.write('trainer 0 stdout: %s\n' % pickle.loads(tr0_out))
+        with open("/tmp/tr0_err.log", "r") as fn:
+            sys.stderr.write('trainer 0 stderr: %s\n' % fn.read())
+        if stat1 == 0:
+            sys.stderr.write('trainer 1 stdout: %s\n' % pickle.loads(tr1_out))
+        with open("/tmp/tr1_err.log", "r") as fn:
+            sys.stderr.write('trainer 1 stderr: %s\n' % fn.read())
 
         return pickle.loads(tr0_out), pickle.loads(tr1_out)
 
@@ -403,13 +430,13 @@ class TestDistBase(unittest.TestCase):
         worker_endpoints = self._ps_endpoints.split(",")
         w0_ep, w1_ep = worker_endpoints
 
-        tr_cmd = "%s %s --role trainer --endpoints %s --trainer_id %d --current_endpoint %s --update_method nccl2"
+        tr_cmd = "%s %s --role trainer --endpoints %s --trainer_id %d --current_endpoint %s --update_method nccl2 --lr %f"
         tr0_cmd = tr_cmd % \
                   (self._python_interp, model, self._ps_endpoints,
-                   0, w0_ep)
+                   0, w0_ep, self._lr / 2)
         tr1_cmd = tr_cmd % \
                   (self._python_interp, model, self._ps_endpoints,
-                   1, w1_ep)
+                   1, w1_ep, self._lr / 2)
 
         if self._mem_opt:
             tr0_cmd += " --mem_opt"
@@ -474,6 +501,7 @@ class TestDistBase(unittest.TestCase):
             "PYTHONPATH": os.getenv("PYTHONPATH", ""),
             "LD_LIBRARY_PATH": os.getenv("LD_LIBRARY_PATH", ""),
             "FLAGS_fraction_of_gpu_memory_to_use": "0.15",
+            "FLAGS_rpc_deadline": "5000",  # 5sec to fail fast
             "FLAGS_cudnn_deterministic": "1",
             "http_proxy": "",
             "NCCL_P2P_DISABLE": "1"

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

@@ -36,7 +36,7 @@ class TestDistMnistNCCL2(TestDistBase):
     def test_dist_train(self):
         import paddle.fluid as fluid
         if fluid.core.is_compiled_with_cuda():
-            self.check_with_place("dist_mnist.py", delta=1)
+            self.check_with_place("dist_mnist.py", delta=1e-5)
 
 
 class TestDistMnist2x2Lars(TestDistBase):

python/paddle/fluid/tests/unittests/test_imperative.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 sys
+import numpy as np
+
+import paddle.fluid as fluid
+from paddle.fluid import core
+
+
+class MyLayer(fluid.imperative.PyLayer):
+    def __init__(self):
+        super(MyLayer, self).__init__()
+
+    def forward(self, inputs):
+        x = fluid.layers.relu(inputs[0])
+        self._x_for_debug = x
+        return [fluid.layers.elementwise_mul(x, x)]
+
+
+class TestImperative(unittest.TestCase):
+    def test_layer(self):
+        with fluid.imperative.guard():
+            cl = core.Layer()
+            cl.forward([])
+            l = fluid.imperative.PyLayer()
+            l.forward([])
+
+    def test_layer_in_out(self):
+        with fluid.imperative.guard():
+            l = MyLayer()
+            x = l(np.array([1.0, 2.0, -1.0], dtype=np.float32))[0]
+            self.assertIsNotNone(x)
+            sys.stderr.write("%s output: %s\n" % (x, x._numpy()))
+            x._backward()
+            sys.stderr.write("grad %s\n" % l._x_for_debug._gradient())
+
+
+if __name__ == '__main__':
+    unittest.main()

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

@@ -846,6 +846,15 @@ class TestBook(unittest.TestCase):
             out = layers.cross_entropy(x, label, False, 4)
             self.assertIsNotNone(out)
 
+    def test_bpr_loss(self):
+        program = Program()
+        with program_guard(program):
+            x = layers.data(name="x", shape=[30, 10], dtype="float32")
+            label = layers.data(name="label", shape=[30, 1], dtype="int32")
+            out = layers.bpr_loss(x, label)
+            self.assertIsNotNone(out)
+        print(str(program))
+
     def test_expand(self):
         program = Program()
         with program_guard(program):

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

@@ -22,6 +22,15 @@ from paddle.fluid.framework import Program, program_guard
 from paddle.fluid.transpiler import memory_optimize
 
 
+def _get_vars(prog):
+    assert (isinstance(prog, Program))
+    all_vars = set()
+    for op in prog.global_block().ops:
+        all_vars.update(op.input_arg_names)
+        all_vars.update(op.output_arg_names)
+    return all_vars
+
+
 class TestControlFlowGraph(unittest.TestCase):
     def setUp(self):
         program = Program()
@@ -37,11 +46,11 @@ class TestControlFlowGraph(unittest.TestCase):
         self.program = program
 
     def test_control_flow_graph(self):
-        print("before optimization")
-        print(str(self.program))
-        result_program = memory_optimize(self.program)
-        print("after optimization")
-        print(str(result_program))
+        result_program = self.program.clone()
+        memory_optimize(self.program)
+        old_vars = _get_vars(self.program)
+        new_vars = _get_vars(result_program)
+        self.assertTrue(old_vars != new_vars)
 
 
 class TestMemoryTranspiler2(unittest.TestCase):
@@ -58,14 +67,22 @@ class TestMemoryTranspiler2(unittest.TestCase):
             avg_cost = layers.mean(cost)
             opt = optimizer.SGD(learning_rate=0.001)
             opt.minimize(avg_cost)
+        self.skip_set = set([cost.name, fc.name])
         self.program = program
 
     def test_inplace_ops(self):
-        print("before optimization")
-        print(str(self.program))
-        result_program = memory_optimize(self.program)
-        print("after optimization")
-        print(str(result_program))
+        result_program = self.program.clone()
+        memory_optimize(self.program)
+        old_vars = _get_vars(self.program)
+        new_vars = _get_vars(result_program)
+        self.assertTrue(old_vars != new_vars)
+
+    def test_skip_opt(self):
+        result_program = self.program.clone()
+        memory_optimize(self.program, skip_opt_set=self.skip_set)
+        old_vars = _get_vars(self.program)
+        new_vars = _get_vars(result_program)
+        self.assertTrue(old_vars != new_vars)
 
 
 class TestMemoryTranspiler3(unittest.TestCase):

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

@@ -63,6 +63,7 @@ class TestSpliteSelectedRows(unittest.TestCase):
         # expected output selected rows
         expected_out0_rows = [0, 4]
         expected_out1_rows = [0, 2]
+        expected_out2_rows = []
         expected_out4_rows = [0]
 
         op = Operator(
@@ -75,6 +76,7 @@ class TestSpliteSelectedRows(unittest.TestCase):
 
         self.assertEqual(outs[0].rows(), expected_out0_rows)
         self.assertEqual(outs[1].rows(), expected_out1_rows)
+        self.assertEqual(outs[2].rows(), expected_out2_rows)
         self.assertEqual(outs[4].rows(), expected_out4_rows)
 
         self.assertEqual(outs[0].height(), height_sections[0])
@@ -84,6 +86,9 @@ class TestSpliteSelectedRows(unittest.TestCase):
         self.assertAlmostEqual(4.0, np.array(outs[1].get_tensor())[1, 1])
         self.assertAlmostEqual(8.0, np.array(outs[4].get_tensor())[0, 1])
 
+        self.assertEqual(outs[2].numel(), 0)
+        self.assertEqual(outs[3].numel(), 0)
+
     def check_grad_with_place(self, place):
         scope = core.Scope()
         height = 10

python/paddle/fluid/transpiler/memory_optimization_transpiler.py

@@ -14,6 +14,7 @@
 
 from __future__ import print_function
 
+import six
 from collections import defaultdict, MutableSet
 from .. import core
 from ... import compat as cpt
@@ -470,8 +471,21 @@ def memory_optimize(input_program,
     Returns:
         None
     """
+
+    def to_name_str(var):
+        if isinstance(var, Variable):
+            return var.desc.name()
+        elif isinstance(var, str):
+            return var
+        elif isinstance(var, six.string_types):
+            return str(var)
+        else:
+            raise TypeError(str(var) + " should be Variable or str")
+
     if level != 0 and level != 1:
         raise ValueError("only support opt_level 0 or 1.")
+    if skip_opt_set is not None and not isinstance(skip_opt_set, set):
+        raise ValueError("only support skip_opt_set as set.")
     global PRINT_LOG
     PRINT_LOG = print_log
     if skip_grads:
@@ -486,6 +500,8 @@ def memory_optimize(input_program,
             skip_opt_set = grad_set
         else:
             skip_opt_set.update(grad_set)
+    if skip_opt_set is not None:
+        skip_opt_set = set(map(to_name_str, skip_opt_set))
     cfgs = _get_cfgs(input_program)
     for cfg in cfgs:
         cfg.memory_optimize(skip_opt_set=skip_opt_set, level=level)

python/setup.py.in

@@ -101,6 +101,7 @@ packages=['paddle',
           'paddle.dataset',
           'paddle.reader',
           'paddle.fluid',
+          'paddle.fluid.imperative',
           'paddle.fluid.proto',
           'paddle.fluid.proto.profiler',
           'paddle.fluid.layers',

tools/print_signatures.py

@@ -27,6 +27,8 @@ import pydoc
 
 member_dict = collections.OrderedDict()
 
+experimental_namespace = {"paddle.fluid.imperative"}
+
 
 def visit_member(parent_name, member):
     cur_name = ".".join([parent_name, member.__name__])
@@ -51,6 +53,8 @@ def visit_member(parent_name, member):
 
 
 def visit_all_module(mod):
+    if (mod.__name__ in experimental_namespace):
+        return
     for member_name in (
             name
             for name in (mod.__all__ if hasattr(mod, "__all__") else dir(mod))