Merge branch 'develop' into mkldnn_test

paddle/fluid/API.spec

@@ -75,7 +75,8 @@ paddle.fluid.layers.conv2d_transpose ArgSpec(args=['input', 'num_filters', 'outp
 paddle.fluid.layers.conv3d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None))
 paddle.fluid.layers.sequence_expand ArgSpec(args=['x', 'y', 'ref_level', 'name'], varargs=None, keywords=None, defaults=(-1, None))
 paddle.fluid.layers.sequence_expand_as ArgSpec(args=['x', 'y', 'name'], varargs=None, keywords=None, defaults=(None,))
-paddle.fluid.layers.sequence_pad ArgSpec(args=['x', 'pad_value', 'maxlen'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.sequence_pad ArgSpec(args=['x', 'pad_value', 'maxlen', 'name'], varargs=None, keywords=None, defaults=(None, None))
+paddle.fluid.layers.sequence_unpad ArgSpec(args=['x', 'length', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.lstm_unit ArgSpec(args=['x_t', 'hidden_t_prev', 'cell_t_prev', 'forget_bias', 'param_attr', 'bias_attr', 'name'], varargs=None, keywords=None, defaults=(0.0, None, None, None))
 paddle.fluid.layers.reduce_sum ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None))
 paddle.fluid.layers.reduce_mean ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None))
@@ -84,6 +85,7 @@ paddle.fluid.layers.reduce_min ArgSpec(args=['input', 'dim', 'keep_dim', 'name']
 paddle.fluid.layers.reduce_prod ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None))
 paddle.fluid.layers.sequence_first_step ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.sequence_last_step ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.layers.sequence_slice ArgSpec(args=['input', 'offset', 'length', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.dropout ArgSpec(args=['x', 'dropout_prob', 'is_test', 'seed', 'name'], varargs=None, keywords=None, defaults=(False, None, None))
 paddle.fluid.layers.split ArgSpec(args=['input', 'num_or_sections', 'dim', 'name'], varargs=None, keywords=None, defaults=(-1, None))
 paddle.fluid.layers.ctc_greedy_decoder ArgSpec(args=['input', 'blank', 'name'], varargs=None, keywords=None, defaults=(None,))

paddle/fluid/framework/details/op_handle_base.h

@@ -64,7 +64,8 @@ class OpHandleBase {
   virtual bool IsMultiDeviceTransfer() { return false; }
 
   const platform::DeviceContext *DeviceContext(platform::Place place) {
-    return dev_ctxes_[place];
+    auto it = dev_ctxes_.find(place);
+    return it != dev_ctxes_.end() ? it->second : nullptr;
   }
 
   void SetDeviceContext(platform::Place place, platform::DeviceContext *ctx_) {

paddle/fluid/framework/executor.cc

@@ -46,6 +46,41 @@ ExecutorPrepareContext::~ExecutorPrepareContext() {
   VLOG(5) << "destroy ExecutorPrepareContext";
 }
 
+template <typename RefCntMap>
+static void DeleteUnusedTensors(const Scope& scope, const OperatorBase* op,
+                                GarbageCollector<Tensor>* gc,
+                                RefCntMap* ref_cnts) {
+  std::unordered_set<Tensor*> erase_tensors;
+
+  auto handler = [&](const VariableNameMap& name_map) {
+    for (auto& name_pair : name_map) {
+      for (auto& name : name_pair.second) {
+        auto it = ref_cnts->find(name);
+        if (it == ref_cnts->end()) continue;
+        if ((it->second)-- == 1) {
+          auto* var = scope.FindVar(name);
+          if (var != nullptr) {
+            VLOG(10) << "Erase tensor \'" << name << "\'";
+            if (var->IsType<LoDTensor>()) {
+              erase_tensors.insert(var->GetMutable<LoDTensor>());
+            } else if (var->IsType<SelectedRows>()) {
+              erase_tensors.insert(
+                  var->GetMutable<SelectedRows>()->mutable_value());
+            }
+          }
+        }
+      }
+    }
+  };
+
+  handler(op->Inputs());
+  handler(op->Outputs());
+
+  if (!erase_tensors.empty()) {
+    gc->Add(erase_tensors);
+  }
+}
+
 Executor::Executor(const platform::Place& place) : place_(place) {}
 
 void Executor::Close() {
@@ -66,7 +101,7 @@ void InitializeVariable(Variable* var, proto::VarType::Type var_type) {
   } else if (var_type == proto::VarType::FETCH_LIST) {
     var->GetMutable<FeedFetchList>();
   } else if (var_type == proto::VarType::STEP_SCOPES) {
-    var->GetMutable<std::vector<framework::Scope>>();
+    var->GetMutable<std::vector<framework::Scope*>>();
   } else if (var_type == proto::VarType::LOD_RANK_TABLE) {
     var->GetMutable<LoDRankTable>();
   } else if (var_type == proto::VarType::LOD_TENSOR_ARRAY) {
@@ -331,9 +366,13 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
   }
 
   int64_t max_memory_size = GetEagerDeletionThreshold();
-
   std::unique_ptr<GarbageCollector<Tensor>> gc;
-  if (max_memory_size >= 0) {
+  // WhileOp would set keep_kids to false
+  // WhileGradOp would need the scopes created in WhileOp
+  // Perhaps, we should not perform eager deletion in WhileOp
+  // The scopes and variables created by WhileOp would be deleted
+  // in WhileGradOp.
+  if (max_memory_size >= 0 && !keep_kids) {
     ctx->ResetReferenceCount();
 #ifdef PADDLE_WITH_CUDA
     if (platform::is_gpu_place(place_)) {
@@ -352,45 +391,8 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
     op->Run(*local_scope, place_);
 
     if (gc != nullptr) {
-      std::vector<std::string> erase_vars;
-      for (auto& input : op->Inputs()) {
-        for (auto& input_name : input.second) {
-          auto it = ctx->cur_ref_cnts_.find(input_name);
-          if (it == ctx->cur_ref_cnts_.end()) continue;
-          if (it->second == 1) {  // should delete it
-            erase_vars.emplace_back(input_name);
-            ctx->cur_ref_cnts_.erase(input_name);
-          } else {
-            --(it->second);
-          }
-        }
-      }
-
-      for (auto& output : op->Outputs()) {
-        for (auto& output_name : output.second) {
-          auto it = ctx->cur_ref_cnts_.find(output_name);
-          if (it == ctx->cur_ref_cnts_.end()) continue;
-          if (it->second == 1) {
-            erase_vars.emplace_back(output_name);
-            ctx->cur_ref_cnts_.erase(output_name);
-          } else {
-            --(it->second);
-          }
-        }
-      }
-
-      if (!erase_vars.empty()) {
-        std::vector<framework::LoDTensor*> erase_tensors;
-        for (auto& name : erase_vars) {
-          auto* var = local_scope->FindVar(name);
-          if (var == nullptr) continue;
-          if (var->IsType<framework::LoDTensor>()) {
-            auto* tensor = var->GetMutable<framework::LoDTensor>();
-            erase_tensors.push_back(tensor);
-          }
-        }
-        if (!erase_tensors.empty()) gc->Add(erase_tensors);
-      }
+      DeleteUnusedTensors(*local_scope, op.get(), gc.get(),
+                          &(ctx->cur_ref_cnts_));
     }
 
     if (FLAGS_benchmark) {

paddle/fluid/framework/executor.h

@@ -32,38 +32,32 @@ template <typename T>
 std::unordered_map<std::string, T> GetNonPersistableReferenceCount(
     const ProgramDesc& prog, size_t block_id) {
   auto& block = prog.Block(block_id);
-  std::unordered_set<std::string> ignored_vars;
   std::unordered_map<std::string, T> ref_cnts;
 
-  for (auto var_desc : block.AllVars()) {
-    auto type = var_desc->Proto()->type().type();
-    if (type != proto::VarType::LOD_TENSOR || var_desc->Persistable()) {
-      ignored_vars.insert(var_desc->Name());  // ignore persistable vars
-    }
-  }
-
-  for (auto op_desc : block.AllOps()) {
-    for (auto& input : op_desc->Inputs()) {
-      for (auto& input_name : input.second) {
-        if (!ignored_vars.count(input_name)) {
-          if (ref_cnts.count(input_name))
-            ++ref_cnts[input_name];
-          else
-            ref_cnts[input_name] = 1;
+  auto update_ref_cnts = [&](OpDesc* op_desc, const VariableNameMap& name_map) {
+    for (auto& name_pair : name_map) {
+      for (auto& name : name_pair.second) {
+        auto* var_desc = block.FindVar(name);
+        if (var_desc == nullptr || var_desc->Persistable()) continue;
+        auto type = var_desc->Proto()->type().type();
+        if (type != proto::VarType::LOD_TENSOR &&
+            type != proto::VarType::SELECTED_ROWS) {
+          continue;
         }
-      }
-    }
 
-    for (auto& output : op_desc->Outputs()) {
-      for (auto output_name : output.second) {
-        if (!ignored_vars.count(output_name)) {
-          if (ref_cnts.count(output_name))
-            ++ref_cnts[output_name];
-          else
-            ref_cnts[output_name] = 1;
+        auto it = ref_cnts.find(name);
+        if (it != ref_cnts.end()) {
+          ++it->second;
+        } else {
+          ref_cnts[name] = 1;
         }
       }
     }
+  };
+
+  for (auto op_desc : block.AllOps()) {
+    update_ref_cnts(op_desc, op_desc->Inputs());
+    update_ref_cnts(op_desc, op_desc->Outputs());
   }
   return ref_cnts;
 }

paddle/fluid/framework/feed_fetch_method.cc

@@ -27,8 +27,7 @@ void SetFeedVariable(Scope* scope, const LoDTensor& input,
   // be created.
   VLOG(3) << "SetFeedVariable name=" << var_name << " index=" << index;
   Variable* g_feed_value = scope->Var(var_name);
-  auto& feed_inputs =
-      *(g_feed_value->GetMutable<std::vector<paddle::framework::LoDTensor>>());
+  auto& feed_inputs = *(g_feed_value->GetMutable<FeedFetchList>());
   if (index >= feed_inputs.size()) {
     feed_inputs.resize(index + 1);
   }

paddle/fluid/framework/naive_executor.cc

@@ -37,7 +37,7 @@ static void InitializeVariable(Variable *var, proto::VarType::Type var_type) {
   } else if (var_type == proto::VarType::FETCH_LIST) {
     var->GetMutable<FeedFetchList>();
   } else if (var_type == proto::VarType::STEP_SCOPES) {
-    var->GetMutable<std::vector<framework::Scope>>();
+    var->GetMutable<std::vector<framework::Scope *>>();
   } else if (var_type == proto::VarType::LOD_RANK_TABLE) {
     var->GetMutable<LoDRankTable>();
   } else if (var_type == proto::VarType::LOD_TENSOR_ARRAY) {

paddle/fluid/framework/operator.cc

@@ -149,9 +149,17 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
     platform::SetDeviceId(dev_id);
 #endif
   }
-  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
-  platform::RecordEvent record_event(Type(), pool.Get(place));
-  RunImpl(scope, place);
+
+  // The profile has a process-wide mutex, results in serious performance issue
+  // in concurrency scenerio. Here use an `if` to fix this issue.
+  // Please not remove the `if`, ask @Superjomn if there are any concern.
+  if (platform::IsProfileEnabled()) {
+    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+    platform::RecordEvent record_event(Type(), pool.Get(place));
+    RunImpl(scope, place);
+  } else {
+    RunImpl(scope, place);
+  }
   VLOG(3) << place << " " << DebugStringEx(&scope);
 }
 

paddle/fluid/framework/parallel_executor.cc

@@ -307,6 +307,10 @@ ParallelExecutor::~ParallelExecutor() {
       }
     }
   }
+
+  // member_ must be destructed before gcs_ since the destructor of
+  // ReferenceCountOpHandle use raw pointers of gcs_ inside.
+  member_.reset();
 }
 
 }  // namespace framework

paddle/fluid/framework/parallel_executor.h

@@ -75,7 +75,7 @@ class ParallelExecutor {
  private:
   void BCastParamsToDevices(const std::unordered_set<std::string> &vars) const;
 
-  ParallelExecutorPrivate *member_;
+  std::unique_ptr<ParallelExecutorPrivate> member_;
 
 #ifdef PADDLE_WITH_CUDA
   // ref_cnts_ is only initialized when ParallelExecutor constructs, and then

paddle/fluid/framework/scope.cc

@@ -49,18 +49,18 @@ int64_t GetEagerDeletionThreshold() {
 Scope::~Scope() { DropKids(); }
 
 Scope& Scope::NewScope() const {
-  std::unique_lock<std::mutex> lock(mutex_);
+  std::lock_guard<std::mutex> lock(mutex_);
   kids_.push_back(new Scope(this));
   return *kids_.back();
 }
 
 Variable* Scope::Var(const std::string& name) {
-  std::unique_lock<std::mutex> lock(mutex_);
+  std::lock_guard<std::mutex> lock(mutex_);
   return VarInternal(name);
 }
 
 Variable* Scope::Var(std::string* name) {
-  std::unique_lock<std::mutex> lock(mutex_);
+  std::lock_guard<std::mutex> lock(mutex_);
   auto new_name = string::Sprintf("%p.%d", this, vars_.size());
   if (name != nullptr) {
     *name = new_name;
@@ -69,29 +69,34 @@ Variable* Scope::Var(std::string* name) {
 }
 
 Variable* Scope::FindVar(const std::string& name) const {
-  std::unique_lock<std::mutex> lock(mutex_);
+  std::lock_guard<std::mutex> lock(mutex_);
   return FindVarInternal(name);
 }
 
+Variable* Scope::FindLocalVar(const std::string& name) const {
+  std::lock_guard<std::mutex> lock(mutex_);
+  return FindVarLocally(name);
+}
+
 const Scope* Scope::FindScope(const Variable* var) const {
-  std::unique_lock<std::mutex> lock(mutex_);
+  std::lock_guard<std::mutex> lock(mutex_);
   return FindScopeInternal(var);
 }
 
 void Scope::DropKids() {
-  std::unique_lock<std::mutex> lock(mutex_);
+  std::lock_guard<std::mutex> lock(mutex_);
   for (Scope* s : kids_) delete s;
   kids_.clear();
 }
 
 bool Scope::HasKid(const Scope* scope) const {
-  std::unique_lock<std::mutex> lock(mutex_);
+  std::lock_guard<std::mutex> lock(mutex_);
   auto it = std::find(this->kids_.begin(), this->kids_.end(), scope);
   return it != this->kids_.end();
 }
 
 std::vector<std::string> Scope::LocalVarNames() const {
-  std::unique_lock<std::mutex> lock(mutex_);
+  std::lock_guard<std::mutex> lock(mutex_);
   std::vector<std::string> known_vars;
   known_vars.reserve(this->vars_.size());
   for (auto& p : vars_) {
@@ -101,7 +106,7 @@ std::vector<std::string> Scope::LocalVarNames() const {
 }
 
 void Scope::DeleteScope(Scope* scope) const {
-  std::unique_lock<std::mutex> lock(mutex_);
+  std::lock_guard<std::mutex> lock(mutex_);
   auto it = std::find(this->kids_.begin(), this->kids_.end(), scope);
   PADDLE_ENFORCE(it != this->kids_.end(), "Cannot find %p as kid scope", scope);
   this->kids_.erase(it);
@@ -114,7 +119,7 @@ void Scope::DeleteScope(Scope* scope) const {
 }
 
 void Scope::EraseVars(const std::vector<std::string>& var_names) {
-  std::unique_lock<std::mutex> lock(mutex_);
+  std::lock_guard<std::mutex> lock(mutex_);
   std::set<std::string> var_set(var_names.begin(), var_names.end());
   for (auto it = vars_.begin(); it != vars_.end();) {
     if (var_set.find(it->first) != var_set.end()) {
@@ -127,12 +132,12 @@ void Scope::EraseVars(const std::vector<std::string>& var_names) {
 
 void Scope::Rename(const std::string& origin_name,
                    const std::string& new_name) const {
-  std::unique_lock<std::mutex> lock(mutex_);
+  std::lock_guard<std::mutex> lock(mutex_);
   RenameInternal(origin_name, new_name);
 }
 
 std::string Scope::Rename(const std::string& origin_name) const {
-  std::unique_lock<std::mutex> lock(mutex_);
+  std::lock_guard<std::mutex> lock(mutex_);
   auto new_name = string::Sprintf("%p.%d", this, vars_.size());
   RenameInternal(origin_name, new_name);
   return new_name;

paddle/fluid/framework/scope.h

@@ -63,6 +63,11 @@ class Scope {
   /// Caller doesn't own the returned Variable.
   Variable* FindVar(const std::string& name) const;
 
+  /// Find a variable in the current scope.
+  /// Return nullptr if cannot find.
+  /// Caller doesn't own the returned Variable.
+  Variable* FindLocalVar(const std::string& name) const;
+
   const Scope* parent() const { return parent_; }
 
   /// Find the scope or an ancestor scope that contains the given variable.

paddle/fluid/framework/tensor_util.cc

@@ -36,6 +36,11 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
   auto size = src.numel() * SizeOfType(src.type());
 
   if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
+    if (src_ptr == dst_ptr) {
+      VLOG(3) << "Skip copy the same data async from " << src_place << " to "
+              << dst_place;
+      return;
+    }
     memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
                  boost::get<platform::CPUPlace>(src_place), src_ptr, size);
   }
@@ -71,6 +76,11 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
     auto stream =
         reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream();
     if (platform::is_same_place(src_place, dst_place)) {
+      if (src_ptr == dst_ptr) {
+        VLOG(3) << "Skip copy the same data async from " << src_place << " to "
+                << dst_place;
+        return;
+      }
       memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size,
                    stream);
     } else {
@@ -114,6 +124,11 @@ void TensorCopySync(const Tensor& src, const platform::Place& dst_place,
   auto dst_ptr = dst->mutable_data(dst_place, src.type());
   auto size = src.numel() * SizeOfType(src.type());
   if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
+    if (src_ptr == dst_ptr) {
+      VLOG(3) << "Skip copy the same data from " << src_place << " to "
+              << dst_place;
+      return;
+    }
     memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
                  boost::get<platform::CPUPlace>(src_place), src_ptr, size);
   }
@@ -130,6 +145,11 @@ void TensorCopySync(const Tensor& src, const platform::Place& dst_place,
     memory::Copy(dst_gpu_place, dst_ptr, src_cpu_place, src_ptr, size, nullptr);
   } else if (platform::is_gpu_place(src_place) &&
              platform::is_gpu_place(dst_place)) {
+    if (src_ptr == dst_ptr && platform::is_same_place(src_place, dst_place)) {
+      VLOG(3) << "Skip copy the same data from " << src_place << " to "
+              << dst_place;
+      return;
+    }
     auto src_gpu_place = boost::get<platform::CUDAPlace>(src_place);
     auto dst_gpu_place = boost::get<platform::CUDAPlace>(dst_place);
     memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, nullptr);

paddle/fluid/framework/tensor_util_test.cc

@@ -41,6 +41,11 @@ TEST(TensorCopy, Tensor) {
     EXPECT_EQ(src_ptr[i], dst_ptr[i]);
   }
 
+  TensorCopy(dst_tensor, *cpu_place, &dst_tensor);
+  for (size_t i = 0; i < 9; ++i) {
+    EXPECT_EQ(src_ptr[i], dst_ptr[i]);
+  }
+
   EXPECT_TRUE(dst_tensor.layout() == src_tensor.layout());
 
   Tensor slice_tensor = src_tensor.Slice(1, 2);
@@ -82,6 +87,15 @@ TEST(TensorCopy, Tensor) {
       EXPECT_EQ(src_ptr[i], dst_ptr[i]);
     }
 
+    // Copy the same tensor
+    TensorCopy(gpu_tensor, *gpu_place, gpu_ctx, &gpu_tensor);
+    gpu_ctx.Wait();
+    const int* dst_ptr_tmp = dst_tensor.data<int>();
+    EXPECT_NE(src_ptr, dst_ptr_tmp);
+    for (size_t i = 0; i < 9; ++i) {
+      EXPECT_EQ(src_ptr[i], dst_ptr_tmp[i]);
+    }
+
     Tensor slice_tensor = src_tensor.Slice(1, 2);
 
     // CPU Slice Tensor to GPU Tensor

paddle/fluid/framework/var_desc.h

@@ -59,6 +59,7 @@ class VarDesc {
  public:
   explicit VarDesc(const std::string &name) {
     desc_.set_name(name);
+    // TODO(paddle-dev): Why default to lodtensor.
     desc_.mutable_type()->set_type(proto::VarType::LOD_TENSOR);
   }
 

paddle/fluid/framework/variable.h

@@ -38,8 +38,12 @@ class Variable {
 
   template <typename T>
   T* GetMutable() {
-    if (!IsType<T>()) {
+    if (!holder_) {
       holder_.reset(new PlaceholderImpl<T>(new T()));
+    } else {
+      PADDLE_ENFORCE(IsType<T>(),
+                     "Variable must be type %s, the holding type is %s",
+                     typeid(T).name(), holder_->Type().name());
     }
     return static_cast<T*>(holder_->Ptr());
   }

paddle/fluid/framework/variable_test.cc

@@ -33,9 +33,10 @@ TEST(Variable, GetMutable) {
   const Tensor& tt = v->Get<Tensor>();
   EXPECT_EQ(1234, tt.content_);
 
-  std::string* s = v->GetMutable<std::string>();
-  *s = "hello";
-
-  const std::string& ss = v->Get<std::string>();
-  EXPECT_EQ("hello", ss);
+  try {
+    v->GetMutable<std::string>();
+  } catch (std::exception& e) {
+    return;
+  }
+  EXPECT_TRUE(false);
 }

paddle/fluid/inference/api/analysis_predictor.cc

@@ -25,9 +25,11 @@
 #include "paddle/fluid/inference/api/paddle_inference_api.h"
 #include "paddle/fluid/inference/api/paddle_inference_pass.h"
 #include "paddle/fluid/inference/utils/singleton.h"
+#include "paddle/fluid/platform/cpu_helper.h"
 #include "paddle/fluid/platform/profiler.h"
 
 DECLARE_bool(profile);
+DECLARE_int32(paddle_num_threads);
 
 namespace paddle {
 
@@ -47,6 +49,9 @@ bool AnalysisPredictor::Init(
   }
 #endif
 
+  // no matter with or without MKLDNN
+  paddle::platform::SetNumThreads(FLAGS_paddle_num_threads);
+
   if (config_.use_gpu) {
     place_ = paddle::platform::CUDAPlace(config_.device);
     LOG(WARNING) << "ir optimize only supports CPU currently, enable_ir_optim "
@@ -335,6 +340,19 @@ bool AnalysisPredictor::LoadProgramDesc() {
   }
   return true;
 }
+
+AnalysisPredictor::~AnalysisPredictor() {
+#if !defined(_WIN32)
+  if (FLAGS_profile) {
+    platform::DisableProfiler(platform::EventSortingKey::kTotal,
+                              "./profile.log");
+  }
+#endif
+  if (sub_scope_) {
+    scope_->DeleteScope(sub_scope_);
+  }
+}
+
 std::unique_ptr<PaddlePredictor> AnalysisPredictor::Clone() {
   auto *x = new AnalysisPredictor(config_);
   x->Init(scope_, inference_program_);

paddle/fluid/inference/api/analysis_predictor.h

@@ -72,6 +72,7 @@ class AnalysisPredictor : public PaddlePredictor {
   template <typename T>
   void GetFetchOne(const framework::LoDTensor &fetchs,
                    PaddleTensor *output_data);
+  ~AnalysisPredictor();
 
  private:
   contrib::AnalysisConfig config_;

paddle/fluid/inference/api/api_impl.cc

@@ -23,9 +23,11 @@ limitations under the License. */
 #include "paddle/fluid/framework/feed_fetch_method.h"
 #include "paddle/fluid/inference/api/api_impl.h"
 #include "paddle/fluid/inference/api/helper.h"
+#include "paddle/fluid/platform/cpu_helper.h"
 #include "paddle/fluid/platform/profiler.h"
 
 DEFINE_bool(profile, false, "Turn on profiler for fluid");
+DECLARE_int32(paddle_num_threads);
 
 namespace paddle {
 namespace {
@@ -72,6 +74,9 @@ bool NativePaddlePredictor::Init(
   }
 #endif
 
+  // no matter with or without MKLDNN
+  paddle::platform::SetNumThreads(FLAGS_paddle_num_threads);
+
   if (config_.use_gpu) {
     place_ = paddle::platform::CUDAPlace(config_.device);
   } else {

paddle/fluid/operators/CMakeLists.txt

@@ -300,7 +300,7 @@ op_library(flatten_op DEPS reshape_op)
 op_library(sequence_pad_op DEPS sequence_padding)
 op_library(unstack_op DEPS stack_op)
 op_library(fake_quantize_op DEPS memory)
-op_library(fusion_lstm_op DEPS cpu_lstm_compute)
+op_library(fusion_lstm_op DEPS jit_kernel)
 if (WITH_GPU)
     op_library(conv_op DEPS vol2col depthwise_conv im2col)
     op_library(layer_norm_op DEPS cub)

paddle/fluid/operators/adam_op.h

@@ -18,6 +18,7 @@ limitations under the License. */
 #include <vector>
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/detail/safe_ref.h"
+#include "paddle/fluid/operators/math/algorithm.h"
 #include "paddle/fluid/operators/math/selected_rows_functor.h"
 #include "paddle/fluid/platform/for_range.h"
 
@@ -199,23 +200,9 @@ struct SparseAdamFunctor {
         row_numel_(row_numel),
         row_count_(row_count) {}
 
-  inline HOSTDEVICE int64_t BinarySearchInRows(int64_t row) const {
-    int64_t beg = 0, end = row_count_ - 1;
-    while (beg <= end) {
-      auto mid = ((beg + end) >> 1);
-      if (rows_[mid] == row)
-        return mid;
-      else if (rows_[mid] < row)
-        beg = mid + 1;
-      else
-        end = mid - 1;
-    }
-    return -1;
-  }
-
   inline HOSTDEVICE void operator()(size_t i) const {
-    int64_t row = i / row_numel_;
-    auto row_idx = BinarySearchInRows(row);
+    auto row_idx =
+        math::BinarySearch<int64_t>(rows_, row_count_, i / row_numel_);
     T g = row_idx >= 0 ? grad_[row_idx * row_numel_ + i % row_numel_] : 0;
 
     // The following code is the same as dense

paddle/fluid/operators/fill_constant_op.cc

@@ -70,6 +70,12 @@ class FillConstantOp : public framework::OperatorBase {
   }
 };
 
+class FillConstantOpVarTypeInference : public framework::VarTypeInference {
+ public:
+  void operator()(const framework::OpDesc &op_desc,
+                  framework::BlockDesc *block) const override {}
+};
+
 class FillConstantOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
@@ -102,4 +108,5 @@ Fill up a variable with specified constant value.
 namespace ops = paddle::operators;
 REGISTER_OPERATOR(fill_constant, ops::FillConstantOp,
                   ops::FillConstantInferShape, ops::FillConstantOpMaker,
-                  paddle::framework::EmptyGradOpMaker);
+                  paddle::framework::EmptyGradOpMaker,
+                  ops::FillConstantOpVarTypeInference);

paddle/fluid/operators/isfinite_op.cc

@@ -60,7 +60,7 @@ class OverflowOpMaker : public framework::OpProtoAndCheckerMaker {
               "(Tensor) 1-dim tensor, contains a bool scalar. The output "
               "tensor of overflow operator.");
     AddComment(string::Sprintf(R"DOC(
-Overflow operator.
+Overflow %s operator.
 
 $$Out = any(X)$$
 
@@ -69,6 +69,8 @@ Out = Inf if any X contains Inf,
 Out = Nan if any X contains Nan,
 Out = 0 if no Inf/Nan detected.
 If X contains both Inf/Nan, it will return the first indicator it meeted.
+
+%s
 )DOC",
                                GetName(), GetComments()));
   }

paddle/fluid/operators/math/CMakeLists.txt

@@ -45,8 +45,6 @@ math_library(im2col)
 if (NOT WIN32) # windows do not support avx functions yet.
 math_library(gru_compute DEPS activation_functions math_function)
 math_library(lstm_compute DEPS activation_functions)
-# TODO(TJ): ugly workaround, clean me
-cc_library(cpu_lstm_compute SRCS cpu_lstm_compute.cc DEPS activation_functions cblas cpu_info)
 endif (NOT WIN32)
 
 cc_library(blas SRCS blas.cc DEPS cblas framework_proto device_context)
@@ -76,3 +74,7 @@ if(WITH_GPU)
 endif()
 cc_test(concat_test SRCS concat_test.cc DEPS concat)
 cc_test(cpu_vec_test SRCS cpu_vec_test.cc DEPS blas cpu_info)
+cc_library(jit_kernel 
+    SRCS jit_kernel.cc jit_kernel_blas.cc jit_kernel_exp.cc jit_kernel_lstm.cc
+    DEPS cpu_info cblas activation_functions)
+cc_test(jit_kernel_test SRCS jit_kernel_test.cc DEPS jit_kernel)