Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into port_python3_syntax

Author: velconia

AUTHORS.md

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

benchmark/fluid/fluid_benchmark.py

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

cmake/external/grpc.cmake

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

cmake/generic.cmake

@@ -263,7 +263,7 @@ function(cc_test TARGET_NAME)
              COMMAND ${TARGET_NAME} ${cc_test_ARGS}
              WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
     if (${cc_test_SERIAL})
-        set_property(TEST ${TARGET_NAME} PROPERTY SERIAL 1)
+        set_property(TEST ${TARGET_NAME} PROPERTY RUN_SERIAL 1)
     set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true)
     endif()
   endif()
@@ -328,7 +328,7 @@ function(nv_test TARGET_NAME)
     add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
     add_test(${TARGET_NAME} ${TARGET_NAME})
     if (nv_test_SERIAL)
-        set_property(TEST ${TARGET_NAME} PROPERTY SERIAL 1)
+        set_property(TEST ${TARGET_NAME} PROPERTY RUN_SERIAL 1)
     set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true)
     endif()
   endif()

cmake/inference_lib.cmake

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

doc/fluid/design/ir/draft.md

@@ -64,13 +64,96 @@ can also contain other things that describe some properties of
 the `Graph` or `Graph` nodes. `Attribute` can be passed
 across `Pass`. However, it should be used with care.
 
+```cpp
+class Graph {
+ public:
+  explicit Graph(const ProgramDesc &program);
+
+  bool Has(const std::string &attr_name) const;
+
+  template <typename AttrType>
+  AttrType &Get(const std::string &attr_name) const;
+
+  template <typename AttrType>
+  void Set(const std::string &attr_name, AttrType *attr);
+  const std::unordered_set<ir::Node *> &Nodes() const;
+
+  // Create a normal variable with non-null VarDesc.
+  ir::Node *CreateVarNode(VarDesc *var_desc);
+
+  // Create a normal runnable operator with OpDesc.
+  ir::Node *CreateOpNode(OpDesc *op_desc);
+
+  // Create a control dependency var that connects 2 operations. The
+  // var doesn't hold any data. Other than that, it's no different from
+  // other var, considering dependency analysis.
+  ir::Node *CreateControlDepVar();
+
+  // 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);
+
+  // Clear all node information of the graph and return the ownership of the
+  // nodes.
+  std::vector<std::unique_ptr<ir::Node>> ReleaseNodes();
+};
+```
+
 #### Pass
 
 `Pass` represents a transformation of `Graph`. Its input
 is a `Graph` and its output is also a `Graph`. For example,
 a `Pass` can simply print out the `Graph`. A `Pass`
 can also fuse some `Graph`'s `Node`s.
 
+```cpp
+class Pass {
+ public:
+
+  std::unique_ptr<Graph> Apply(std::unique_ptr<Graph> graph) const {
+    // Some correctness check.
+    auto new_graph = ApplyImpl(std::move(graph));
+    // Some correctness check.
+    return new_graph;
+  }
+
+  // Get a reference to the attributed previously set.
+  template <typename AttrType>
+  AttrType &Get(const std::string &attr_name) const;
+
+  // Set a pointer to the attribute. Pass takes ownership of the attribute.
+  template <typename AttrType>
+  void Set(const std::string &attr_name, AttrType *attr) ;
+
+  // Set a pointer to the attribute. Pass doesn't take ownership. Caller
+  // should delete the attribute.
+  template <typename AttrType>
+  void SetNotOwned(const std::string &attr_name, AttrType *attr);
+
+ protected:
+  virtual std::unique_ptr<Graph> ApplyImpl(std::unique_ptr<Graph> graph) const = 0;
+};
+
+// In my_pass.cc
+class MyPass : public Pass {
+ protected:
+  std::unique_ptr<Graph> ApplyImpl(std::unique_ptr<Graph> graph) const override {
+    // do something.
+    return graph;
+  }
+}
+REGISTER_PASS(my_pass, MyPass)
+.RequirePassAttr("places")
+.RequireGraphAttr("dep_vars");
+
+
+// To use the pass.
+auto my_pass = ir::PassRegistry::Instance().Get("my_pass");
+graph = my_pass->Apply(std::move(graph));
+// Note: to force link my_pass.cc, in the code:
+USE_PASS(my_pass);
+```
+
 #### Optimize
 
 `Optimize` contains a series of `Pass` with defined order.
@@ -86,4 +169,17 @@ maintaining the original modeling logic.
 * Graph is transformed from raw model logic to a
 form that is efficient to execute.
 
-Program->ProgramToGraph->Graph->Pass1->Graph->Pass2->Graph->Pass3->Graph->Executor
+```
+// Program->ProgramToGraph->Graph->Pass1->Graph->Pass2->Graph->Pass3->Graph->Executor
+auto graph = Graph(program);
+graph = PassRegistry::Instance().Get("op_fuse_pass").Apply(std::move(grah));
+// For more complex Pass, Optimize Process can provide Pass attributes.
+auto mem_opt_pass = PassRegistry::Instance().Get("memory_optimization_pass");
+mem_opt_pass.SetNotOwned<int>("optimize_level", 1);
+mem_opt_pass->Apply(std::move(graph));
+graph = PassRegistry::Instance().Get("multi_device_pass").Apply(std::move(grah));
+graph = PassRegistry::Instance().Get("multi_device_check_pass").Apply(std::move(grah));
+Executor exe;
+exe.Run(graph);
+
+```

paddle/fluid/framework/CMakeLists.txt

@@ -8,9 +8,9 @@ cc_test(ddim_test SRCS ddim_test.cc DEPS ddim)
 nv_test(dim_test SRCS dim_test.cu DEPS ddim)
 cc_library(data_type SRCS data_type.cc DEPS framework_proto ddim device_context)
 if(WITH_GPU)
-  nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS place memory data_type)
+  nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS place memory data_type device_context)
 else()
-  cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS place memory data_type)
+  cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS place memory data_type device_context)
 endif()
 
 cc_test(tensor_test SRCS tensor_test.cc DEPS tensor)
@@ -99,7 +99,7 @@ else()
 endif()
 
 
-cc_library(parallel_executor SRCS parallel_executor.cc DEPS ssa_graph_builder_factory threaded_ssa_graph_executor scope_buffered_ssa_graph_executor graph)
+cc_library(parallel_executor SRCS parallel_executor.cc DEPS threaded_ssa_graph_executor scope_buffered_ssa_graph_executor graph graph_viz_pass multi_devices_graph_builder ssa_graph_printer ssa_graph_checker)
 
 cc_library(prune SRCS prune.cc DEPS framework_proto)
 cc_test(prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_context)
@@ -110,7 +110,7 @@ cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows)
 
 cc_test(op_kernel_type_test SRCS op_kernel_type_test.cc DEPS place device_context framework_proto)
 cc_test(cow_ptr_tests SRCS details/cow_ptr_test.cc)
-      
+
 # cc_test(channel_test SRCS channel_test.cc)
 cc_test(tuple_test SRCS tuple_test.cc )
 

paddle/fluid/framework/details/CMakeLists.txt

@@ -31,9 +31,6 @@ cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base s
 cc_library(multi_devices_graph_builder SRCS multi_devices_graph_builder.cc DEPS ssa_graph_builder computation_op_handle
         scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle)
 
-
-cc_library(ssa_graph_builder_factory SRCS ssa_graph_builder_factory.cc DEPS multi_devices_graph_builder ssa_graph_printer ssa_graph_checker)
-
 cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS graph framework_proto)
 cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope
         simple_threadpool device_context)