Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into add-test-for-split-ids-op

Author: jacquesqiao

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/API.spec

@@ -170,6 +170,7 @@ paddle.fluid.layers.mean_iou ArgSpec(args=['input', 'label', 'num_classes'], var
 paddle.fluid.layers.relu ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.log ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.crop ArgSpec(args=['x', 'shape', 'offsets', 'name'], varargs=None, keywords=None, defaults=(None, None, None))
+paddle.fluid.layers.rank_loss ArgSpec(args=['label', 'left', 'right', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
 paddle.fluid.layers.open_recordio_file ArgSpec(args=['filename', 'shapes', 'lod_levels', 'dtypes', 'pass_num', 'for_parallel'], varargs=None, keywords=None, defaults=(1, True))
 paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None))
@@ -201,7 +202,6 @@ paddle.fluid.layers.zeros ArgSpec(args=['shape', 'dtype', 'force_cpu'], varargs=
 paddle.fluid.layers.reverse ArgSpec(args=['x', 'axis'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.While.__init__ ArgSpec(args=['self', 'cond', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.While.block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
-paddle.fluid.layers.While.complete ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.Switch.__init__ ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.Switch.case ArgSpec(args=['self', 'condition'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.Switch.default ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
@@ -225,17 +225,14 @@ paddle.fluid.layers.DynamicRNN.static_input ArgSpec(args=['self', 'x'], varargs=
 paddle.fluid.layers.DynamicRNN.step_input ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.DynamicRNN.update_memory ArgSpec(args=['self', 'ex_mem', 'new_mem'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.StaticRNN.__init__ ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,))
-paddle.fluid.layers.StaticRNN.complete_op ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.StaticRNN.memory ArgSpec(args=['self', 'init', 'shape', 'batch_ref', 'init_value', 'init_batch_dim_idx', 'ref_batch_dim_idx'], varargs=None, keywords=None, defaults=(None, None, None, 0.0, 0, 1))
 paddle.fluid.layers.StaticRNN.output ArgSpec(args=['self'], varargs='outputs', keywords=None, defaults=None)
-paddle.fluid.layers.StaticRNN.parent_block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.StaticRNN.step ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.StaticRNN.step_input ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.StaticRNN.step_output ArgSpec(args=['self', 'o'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.StaticRNN.update_memory ArgSpec(args=['self', 'mem', 'var'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.reorder_lod_tensor_by_rank ArgSpec(args=['x', 'rank_table'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.ParallelDo.__init__ ArgSpec(args=['self', 'places', 'use_nccl', 'name'], varargs=None, keywords=None, defaults=(False, None))
-paddle.fluid.layers.ParallelDo.complete_op ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.ParallelDo.do ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.ParallelDo.get_parameters ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.ParallelDo.parent_block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)

paddle/fluid/framework/CMakeLists.txt

@@ -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)

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)