Merge branch 'develop' of https://github.com/paddlepaddle/paddle into add_trt_plugin

merge develop and fix conflicts

Author: NHZlX

AUTHORS.md

@@ -43,6 +43,7 @@
 | qingqing01 | Qing-Qing Dang |
 | reyoung | Yang Yu |
 | Superjom | Chun-Wei Yan |
+| tensor-tang | Jian Tang |
 | tianbingsz | Tian-Bing Xu |
 | tpatejko | Tomasz Patejko |
 | typhoonzero | Yi Wu |

cmake/inference_lib.cmake

@@ -164,7 +164,7 @@ endif()
 set(module "inference")
 copy(inference_lib DEPS ${inference_deps}
   SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.*
-       ${src_dir}/${module}/api/paddle_inference_api.h
+       ${src_dir}/${module}/api/paddle_*.h
        ${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/paddle_inference_pass.h
   DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module}
 )
@@ -202,10 +202,10 @@ copy(third_party DEPS fluid_lib_dist
   DSTS ${FLUID_INFERENCE_INSTALL_DIR} ${FLUID_INFERENCE_INSTALL_DIR}
 )
 
-# only need libpaddle_fluid.so/a and paddle_inference_api.h for inference-only library
+# only need libpaddle_fluid.so/a and paddle_*.h for inference-only library
 copy(inference_api_lib DEPS fluid_lib_dist
   SRCS ${FLUID_INSTALL_DIR}/paddle/fluid/inference/libpaddle_fluid.*
-       ${FLUID_INSTALL_DIR}/paddle/fluid/inference/paddle_inference_api.h
+       ${FLUID_INSTALL_DIR}/paddle/fluid/inference/paddle_*.h
   DSTS ${FLUID_INFERENCE_INSTALL_DIR}/paddle/lib ${FLUID_INFERENCE_INSTALL_DIR}/paddle/include
 )
 

cmake/tensorrt.cmake

@@ -34,4 +34,5 @@ if(TENSORRT_FOUND)
         "Current TensorRT version is v${TENSORRT_MAJOR_VERSION}. ")
     include_directories(${TENSORRT_INCLUDE_DIR})
     list(APPEND EXTERNAL_LIBS ${TENSORRT_LIBRARY})
+    add_definitions(-DPADDLE_WITH_TENSORRT)
 endif()

paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc

@@ -30,8 +30,8 @@ FastThreadedSSAGraphExecutor::FastThreadedSSAGraphExecutor(
       local_scopes_(local_scopes),
       places_(places),
       graph_(std::move(graph)),
-      pool_(strategy.num_threads_ +
-            1),  // add one more thread for generate op_deps
+      pool_(strategy.num_threads_),
+      prepare_pool_(1),  // add one more thread for generate op_deps
       fetch_ctxs_(places) {
   for (auto &op : ir::FilterByNodeWrapper<OpHandleBase>(*graph_)) {
     int dep = static_cast<int>(op->NotReadyInputSize());
@@ -160,7 +160,7 @@ void FastThreadedSSAGraphExecutor::RunOpAsync(
   });
 }
 void FastThreadedSSAGraphExecutor::PrepareAtomicOpDeps() {
-  atomic_op_deps_ = pool_.enqueue([&] {
+  atomic_op_deps_ = prepare_pool_.enqueue([&] {
     auto *op_deps = new std::unordered_map<OpHandleBase *, std::atomic<int>>;
     for (auto &pair : op_deps_) {
       (*op_deps)[pair.first] = pair.second;

paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h

@@ -46,6 +46,7 @@ class FastThreadedSSAGraphExecutor : public SSAGraphExecutor {
   std::vector<OpHandleBase *> bootstrap_ops_;
 
   ::ThreadPool pool_;
+  ::ThreadPool prepare_pool_;
   platform::DeviceContextPool fetch_ctxs_;
   std::atomic<int> remaining_;
 

paddle/fluid/framework/executor.cc

@@ -359,6 +359,7 @@ std::vector<std::shared_ptr<ExecutorPrepareContext>> Executor::Prepare(
 void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
                                   bool create_local_scope, bool create_vars,
                                   bool keep_kids) {
+  PADDLE_ENFORCE_NOT_NULL(scope);
   Scope* local_scope = scope;
   if (create_vars) {
     if (create_local_scope) {

paddle/fluid/framework/ir/CMakeLists.txt

@@ -5,6 +5,7 @@ file(APPEND ${pass_file} "\#include \"paddle/fluid/framework/ir/pass.h\"\n")
 
 
 # Usage: pass_library(target inference) will append to paddle_inference_pass.h
+unset(INFER_IR_PASSES CACHE) # clear the global variable
 function(pass_library TARGET DEST)
     set(options "")
     set(oneValueArgs "")
@@ -15,10 +16,11 @@ function(pass_library TARGET DEST)
     if (${DEST} STREQUAL "base" OR ${DEST} STREQUAL "inference")
         message(STATUS "add pass ${TARGET} ${DEST}")
         file(APPEND ${pass_file} "USE_PASS(${TARGET});\n")
-        set(PASS_LIBRARY ${TARGET} ${PASS_LIBRARY} PARENT_SCOPE)
+        set(INFER_IR_PASSES ${INFER_IR_PASSES} ${TARGET} CACHE INTERNAL "")
     endif()
 endfunction()
 
+
 cc_library(node SRCS node.cc DEPS proto_desc)
 cc_library(graph SRCS graph.cc DEPS node pretty_log)
 cc_library(graph_helper SRCS graph_helper.cc DEPS graph)

paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc

@@ -91,10 +91,10 @@ void FindWhileOp(Graph* graph) {
 #undef OP_SET_IN
 #undef OP_SET_OUT
 
-  auto* X = graph->RetriveNode(34);
-  auto* LSTMOUT = graph->RetriveNode(81);
-  auto* cell_init = graph->RetriveNode(6);
-  auto* hidden_init = graph->RetriveNode(8);
+  auto* X = graph->RetrieveNode(34);
+  auto* LSTMOUT = graph->RetrieveNode(81);
+  auto* cell_init = graph->RetrieveNode(6);
+  auto* hidden_init = graph->RetrieveNode(8);
 
   auto* lstm_op = graph->CreateOpNode(&op_desc);
   PrepareParameters(graph, param);

paddle/fluid/framework/ir/graph.cc

@@ -84,8 +84,6 @@ void CheckProgram(const ProgramDesc &program) {
 
 Graph::Graph(const ProgramDesc &program) : program_(program) {
   CheckProgram(program_);
-  // Make the nodes id start from 0.
-  Node::ResetId();
   auto var_nodes = InitFromProgram(program_);
   ResolveHazard(var_nodes);
 }

paddle/fluid/framework/ir/graph.h

@@ -116,13 +116,17 @@ class Graph {
   // Create a normal variable with non-null VarDesc.
   ir::Node *CreateVarNode(VarDesc *var_desc) {
     PADDLE_ENFORCE(var_desc);
-    return AddNode(new ir::Node(var_desc));
+    auto *x = AddNode(new ir::Node(var_desc));
+    x->SetId(num_node_created_++);
+    return x;
   }
 
   // Create a normal runnable operator with OpDesc.
   ir::Node *CreateOpNode(OpDesc *op_desc) {
     PADDLE_ENFORCE(op_desc);
-    return AddNode(new ir::Node(op_desc));
+    auto *x = AddNode(new ir::Node(op_desc));
+    x->SetId(num_node_created_++);
+    return x;
   }
 
   // Create a control dependency var that connects 2 operations. The
@@ -132,13 +136,17 @@ class Graph {
     // TODO(panyx0718): control var name should be really unique.
     const std::string name = string::Sprintf(
         "%s@%llu", ir::Node::kControlDepVarName, node_set_.size());
-    return AddNode(new ir::Node(name, ir::Node::Type::kVariable));
+    auto *x = AddNode(new ir::Node(name, ir::Node::Type::kVariable));
+    x->SetId(num_node_created_++);
+    return x;
   }
 
   // A more free style way of creating a graph node. Mostly use for test
   // or "copy" from another node. Avoid using it if possible.
   ir::Node *CreateEmptyNode(const std::string &name, ir::Node::Type type) {
-    return AddNode(new ir::Node(name, type));
+    auto *x = AddNode(new ir::Node(name, type));
+    x->SetId(num_node_created_++);
+    return x;
   }
 
   // Clear all node information of the graph and return the ownership of the
@@ -160,7 +168,7 @@ class Graph {
   }
 
   // NOTE low performance, but simple and secure.
-  Node *RetriveNode(int id) {
+  Node *RetrieveNode(int id) {
     for (auto &node : nodes_) {
       if (node.second->id() == id) {
         return node.second.get();
@@ -169,6 +177,7 @@ class Graph {
     return nullptr;
   }
 
+  const ProgramDesc &program() const { return program_; }
   std::map<std::string, std::vector<ir::Node *>> InitFromProgram(
       const ProgramDesc &program);
 
@@ -190,6 +199,7 @@ class Graph {
   std::map<std::string, std::function<void(void)>> attr_dels_;
   std::map<ir::Node *, std::unique_ptr<ir::Node>> nodes_;
   std::unordered_set<ir::Node *> node_set_;
+  size_t num_node_created_{0};  // help to generate a unique node id.
 };
 
 bool IsControlDepVar(const ir::Node &var);