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

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(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_operator variable_helper)
+      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(WITH_NGRAPH)
   cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op)
 endif()
 

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/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>
@@ -58,9 +57,9 @@ 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,
+  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,
@@ -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 ||
@@ -251,7 +250,7 @@ void FusedOperator::Process() {
   }
 }
 
-void FusedOperator::RunImpl(const Scope& scope,
+void NgraphOperator::RunImpl(const Scope& scope,
                              const platform::Place& place) const {
   op_state ng_op_state = PARTIAL_TEST;
   auto& bdesc = pdesc_.Block(block_);
@@ -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_,
+  NgraphEngine ngraph_engine(scope, place, fused_ops_, var_type_map_,
                              persistables_, fetches_, post_op_inputs_,
                              ng_op_state);
-  ngraph_op.Run(scope, place);
+  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_registry.h

@@ -319,7 +319,7 @@ struct OpKernelRegistrarFunctorEx<PlaceType, false, I,
 #define USE_OP(op_type)   \
   USE_OP_ITSELF(op_type); \
   USE_OP_KERNEL(op_type)
-// clang-format off
+// clang-format on
 
 }  // namespace framework
 }  // 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/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/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/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();

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_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"
@@ -425,13 +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"

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_regularizer.py

@@ -15,7 +15,12 @@
 from __future__ import print_function
 
 import unittest
-
+from functools import partial
+import contextlib
+import numpy as np
+import paddle
+import paddle.fluid.core as core
+import paddle.fluid as fluid
 import paddle.fluid.framework as framework
 import paddle.fluid.optimizer as optimizer
 import paddle.fluid.regularizer as regularizer
@@ -97,5 +102,134 @@ class TestL1DecayRegularizer(unittest.TestCase):
         self.assertEqual(block.ops[-3].type, 'sign')
 
 
+def bow_net(data,
+            label,
+            dict_dim,
+            is_sparse=False,
+            emb_dim=128,
+            hid_dim=128,
+            hid_dim2=96,
+            class_dim=2):
+    """
+    BOW net
+    This model is from https://github.com/PaddlePaddle/models:
+    fluid/PaddleNLP/text_classification/nets.py
+    """
+    emb = fluid.layers.embedding(
+        input=data, is_sparse=is_sparse, size=[dict_dim, emb_dim])
+    bow = fluid.layers.sequence_pool(input=emb, pool_type='sum')
+    bow_tanh = fluid.layers.tanh(bow)
+    fc_1 = fluid.layers.fc(input=bow_tanh, size=hid_dim, act="tanh")
+    fc_2 = fluid.layers.fc(input=fc_1, size=hid_dim2, act="tanh")
+    prediction = fluid.layers.fc(input=[fc_2], size=class_dim, act="softmax")
+    cost = fluid.layers.cross_entropy(input=prediction, label=label)
+    avg_cost = fluid.layers.mean(x=cost)
+
+    return avg_cost
+
+
+class TestRegularizer(unittest.TestCase):
+    def setUp(self):
+        self.word_dict = paddle.dataset.imdb.word_dict()
+        reader = paddle.batch(
+            paddle.dataset.imdb.train(self.word_dict), batch_size=8)()
+        self.train_data = [next(reader) for _ in range(5)]
+
+    def get_places(self):
+        places = [core.CPUPlace()]
+        if core.is_compiled_with_cuda():
+            places.append(core.CUDAPlace(0))
+        return places
+
+    @contextlib.contextmanager
+    def scope_prog_guard(self, main_prog, startup_prog):
+        scope = fluid.core.Scope()
+        with fluid.unique_name.guard():
+            with fluid.scope_guard(scope):
+                with fluid.program_guard(main_prog, startup_prog):
+                    yield
+
+    def run_program(self, place, feed_list):
+        exe = fluid.Executor(place)
+        feeder = fluid.DataFeeder(feed_list=feed_list, place=place)
+        exe.run(fluid.default_startup_program())
+
+        main_prog = fluid.default_main_program()
+        param_list = [var.name for var in main_prog.block(0).all_parameters()]
+
+        param_sum = []
+        for data in self.train_data:
+            out = exe.run(main_prog,
+                          feed=feeder.feed(data),
+                          fetch_list=param_list)
+            p_sum = 0
+            for v in out:
+                p_sum += np.sum(np.abs(v))
+            param_sum.append(p_sum)
+        return param_sum
+
+    def check_l2decay_regularizer(self, place, model):
+        main_prog = fluid.framework.Program()
+        startup_prog = fluid.framework.Program()
+        startup_prog.random_seed = 1
+        with self.scope_prog_guard(
+                main_prog=main_prog, startup_prog=startup_prog):
+            data = fluid.layers.data(
+                name="words", shape=[1], dtype="int64", lod_level=1)
+            label = fluid.layers.data(name="label", shape=[1], dtype="int64")
+
+            avg_cost = model(data, label, len(self.word_dict))
+
+            optimizer = fluid.optimizer.Adagrad(
+                learning_rate=0.1,
+                regularization=fluid.regularizer.L2Decay(1.0))
+            optimizer.minimize(avg_cost)
+            param_sum = self.run_program(place, [data, label])
+        return param_sum
+
+    def check_l2decay(self, place, model):
+        main_prog = fluid.framework.Program()
+        startup_prog = fluid.framework.Program()
+        startup_prog.random_seed = 1
+        with self.scope_prog_guard(
+                main_prog=main_prog, startup_prog=startup_prog):
+            data = fluid.layers.data(
+                name="words", shape=[1], dtype="int64", lod_level=1)
+            label = fluid.layers.data(name="label", shape=[1], dtype="int64")
+
+            avg_cost_l2 = model(data, label, len(self.word_dict))
+
+            param_list = fluid.default_main_program().block(0).all_parameters()
+            para_sum = []
+            for para in param_list:
+                para_mul = fluid.layers.square(x=para)
+                para_sum.append(fluid.layers.reduce_sum(input=para_mul))
+            avg_cost_l2 += fluid.layers.sums(para_sum) * .5
+
+            optimizer = fluid.optimizer.Adagrad(learning_rate=0.1)
+            optimizer.minimize(avg_cost_l2)
+            param_sum = self.run_program(place, [data, label])
+        return param_sum
+
+    def test_l2(self):
+        for place in self.get_places():
+            dense_sparse_p_sum = []
+            for sparse in [True, False]:
+                model = partial(bow_net, is_sparse=sparse)
+                framework_l2 = self.check_l2decay_regularizer(place, model)
+                l2 = self.check_l2decay(place, model)
+                assert len(l2) == len(framework_l2)
+                for i in range(len(l2)):
+                    assert np.isclose(a=framework_l2[i], b=l2[i], rtol=5e-5)
+                dense_sparse_p_sum.append(framework_l2)
+
+            assert len(dense_sparse_p_sum[0]) == len(dense_sparse_p_sum[1])
+            for i in range(len(dense_sparse_p_sum[0])):
+                assert np.isclose(
+                    a=dense_sparse_p_sum[0][i],
+                    b=dense_sparse_p_sum[1][i],
+                    rtol=5e-5)
+
+
 if __name__ == '__main__':
     unittest.main()