Merge branch 'develop' into anakin_bug

Author: luotao1

CMakeLists.txt

@@ -204,12 +204,11 @@ include(external/snappy)    # download snappy
 include(external/snappystream)
 include(external/threadpool)
 
+set(WITH_ANAKIN OFF CACHE STRING "Disable Anakin first, will add it later." FORCE)
 if(WITH_GPU)
     include(cuda)
     include(tensorrt)
     include(external/anakin)
-else()
-  set(WITH_ANAKIN OFF CACHE STRING "Anakin is valid only when GPU is set." FORCE)
 endif()
 
 include(cudnn)              # set cudnn libraries, must before configure

doc/fluid/new_docs/advanced_usage/deploy/native_infer.rst

@@ -4,7 +4,7 @@ Paddle 预测 API
 为了更简单方便的预测部署，Fluid 提供了一套高层 API
 用来隐藏底层不同的优化实现。
 
-`预测库相关代码 <https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/contrib/inference>`__
+`预测库相关代码 <https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/inference/api>`__
 包括
 
 -  头文件 ``paddle_inference_api.h`` 定义了所有的接口
@@ -104,5 +104,5 @@ engine
 ------------
 
 -  `inference
-   demos <https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/contrib/inference/demo>`__
--  `复杂单线程/多线程例子 <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/contrib/inference/test_paddle_inference_api_impl.cc>`__
+   demos <https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/inference/api/demo_ci>`__
+-  `复杂单线程/多线程例子 <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/api/api_impl_tester.cc>`__

paddle/fluid/API.spec

@@ -6,7 +6,7 @@ paddle.fluid.Program.create_block ArgSpec(args=['self', 'parent_idx'], varargs=N
 paddle.fluid.Program.current_block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Program.get_desc ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Program.global_block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
-paddle.fluid.Program.inference_optimize ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.Program.inference_optimize ArgSpec(args=['self', 'export_for_deployment'], varargs=None, keywords=None, defaults=(True,))
 paddle.fluid.Program.list_vars ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Program.optimized_guard ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None)
 paddle.fluid.Program.parse_from_string ArgSpec(args=['binary_str'], varargs=None, keywords=None, defaults=None)
@@ -18,6 +18,9 @@ paddle.fluid.Operator.all_attrs ArgSpec(args=['self'], varargs=None, keywords=No
 paddle.fluid.Operator.attr ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Operator.attr_type ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Operator.block_attr ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.Operator.block_attr_id ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.Operator.blocks_attr ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.Operator.blocks_attr_ids ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Operator.has_attr ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Operator.has_kernel ArgSpec(args=['self', 'op_type'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Operator.input ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None)
@@ -52,7 +55,7 @@ paddle.fluid.Inferencer.__init__ ArgSpec(args=['self', 'infer_func', 'param_path
 paddle.fluid.Inferencer.infer ArgSpec(args=['self', 'inputs', 'return_numpy'], varargs=None, keywords=None, defaults=(True,))
 paddle.fluid.DistributeTranspiler.__init__ ArgSpec(args=['self', 'config'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.DistributeTranspiler.get_pserver_program ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None)
-paddle.fluid.DistributeTranspiler.get_startup_program ArgSpec(args=['self', 'endpoint', 'pserver_program'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.DistributeTranspiler.get_startup_program ArgSpec(args=['self', 'endpoint', 'pserver_program', 'startup_program'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.DistributeTranspiler.get_trainer_program ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.DistributeTranspiler.transpile ArgSpec(args=['self', 'trainer_id', 'program', 'pservers', 'trainers', 'sync_mode'], varargs=None, keywords=None, defaults=(None, '127.0.0.1:6174', 1, True))
 paddle.fluid.InferenceTranspiler.__init__ 
@@ -74,7 +77,7 @@ paddle.fluid.io.save_persistables ArgSpec(args=['executor', 'dirname', 'main_pro
 paddle.fluid.io.load_vars ArgSpec(args=['executor', 'dirname', 'main_program', 'vars', 'predicate', 'filename'], varargs=None, keywords=None, defaults=(None, None, None, None))
 paddle.fluid.io.load_params ArgSpec(args=['executor', 'dirname', 'main_program', 'filename'], varargs=None, keywords=None, defaults=(None, None))
 paddle.fluid.io.load_persistables ArgSpec(args=['executor', 'dirname', 'main_program', 'filename'], varargs=None, keywords=None, defaults=(None, None))
-paddle.fluid.io.save_inference_model ArgSpec(args=['dirname', 'feeded_var_names', 'target_vars', 'executor', 'main_program', 'model_filename', 'params_filename'], varargs=None, keywords=None, defaults=(None, None, None))
+paddle.fluid.io.save_inference_model ArgSpec(args=['dirname', 'feeded_var_names', 'target_vars', 'executor', 'main_program', 'model_filename', 'params_filename', 'export_for_deployment'], varargs=None, keywords=None, defaults=(None, None, None, True))
 paddle.fluid.io.load_inference_model ArgSpec(args=['dirname', 'executor', 'model_filename', 'params_filename'], varargs=None, keywords=None, defaults=(None, None))
 paddle.fluid.io.get_inference_program ArgSpec(args=['target_vars', 'main_program'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.initializer.ConstantInitializer.__init__ ArgSpec(args=['self', 'value', 'force_cpu'], varargs=None, keywords=None, defaults=(0.0, False))
@@ -156,6 +159,8 @@ paddle.fluid.layers.relu ArgSpec(args=['x'], varargs=None, keywords=None, defaul
 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.prelu ArgSpec(args=['x', 'mode', 'param_attr', 'name'], varargs=None, keywords=None, defaults=(None, None))
+paddle.fluid.layers.flatten ArgSpec(args=['x', 'axis', 'name'], varargs=None, keywords=None, defaults=(1, 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))
@@ -324,7 +329,7 @@ paddle.fluid.contrib.BeamSearchDecoder.update_array ArgSpec(args=['self', 'array
 paddle.fluid.contrib.memory_usage ArgSpec(args=['program', 'batch_size'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.transpiler.DistributeTranspiler.__init__ ArgSpec(args=['self', 'config'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.transpiler.DistributeTranspiler.get_pserver_program ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None)
-paddle.fluid.transpiler.DistributeTranspiler.get_startup_program ArgSpec(args=['self', 'endpoint', 'pserver_program'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.transpiler.DistributeTranspiler.get_startup_program ArgSpec(args=['self', 'endpoint', 'pserver_program', 'startup_program'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.transpiler.DistributeTranspiler.get_trainer_program ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.transpiler.DistributeTranspiler.transpile ArgSpec(args=['self', 'trainer_id', 'program', 'pservers', 'trainers', 'sync_mode'], varargs=None, keywords=None, defaults=(None, '127.0.0.1:6174', 1, True))
 paddle.fluid.transpiler.InferenceTranspiler.__init__ 

paddle/fluid/framework/details/exception_holder.h

@@ -14,6 +14,7 @@
 
 #pragma once
 
+#include "glog/logging.h"
 #include "paddle/fluid/platform/enforce.h"
 
 namespace paddle {
@@ -22,55 +23,66 @@ namespace details {
 
 class ExceptionHolder {
  public:
-  void Catch(const platform::EnforceNotMet& exp) {
-    std::lock_guard<std::mutex> lock(mu_);
-    exception_.reset(new platform::EnforceNotMet(exp));
-    type_ = kEnforceNotMet;
-  }
-
-  void Catch(const platform::EOFException& exp) {
-    std::lock_guard<std::mutex> lock(mu_);
-    // EOFException will not cover up existing EnforceNotMet.
-    if (exception_.get() == nullptr) {
-      exception_.reset(new platform::EOFException(exp));
-      type_ = kEOF;
+  void Catch(std::exception_ptr eptr) {
+    try {
+      std::rethrow_exception(eptr);
+    } catch (platform::EOFException exp) {
+      Catch(exp);
+    } catch (platform::EnforceNotMet exp) {
+      Catch(exp);
+    } catch (...) {
+      LOG(FATAL) << "Unknown exception caught";
     }
   }
 
-  bool ExceptionCatched() const {
+  bool IsCaught() const {
     std::lock_guard<std::mutex> lock(mu_);
     return exception_.get() != nullptr;
   }
 
-  void Throw() {
+  void ReThrow() {
     std::lock_guard<std::mutex> lock(mu_);
     switch (type_) {
       case kNone:
         break;
       case kEnforceNotMet: {
         auto e = *static_cast<platform::EnforceNotMet*>(exception_.get());
         throw e;
-        break;
       }
       case kEOF: {
         auto e = *static_cast<platform::EOFException*>(exception_.get());
         throw e;
-        break;
       }
-      default:
-        LOG(FATAL) << "Unknown exception.";
     }
-    exception_.reset();
-    type_ = kNone;
+    ClearImpl();
   }
 
   void Clear() {
     std::lock_guard<std::mutex> lock(mu_);
+    ClearImpl();
+  }
+
+ private:
+  void ClearImpl() {
     exception_.reset();
     type_ = kNone;
   }
 
- private:
+  void Catch(const platform::EnforceNotMet& exp) {
+    std::lock_guard<std::mutex> lock(mu_);
+    exception_.reset(new platform::EnforceNotMet(exp));
+    type_ = kEnforceNotMet;
+  }
+
+  void Catch(const platform::EOFException& exp) {
+    std::lock_guard<std::mutex> lock(mu_);
+    // EOFException will not cover up existing EnforceNotMet.
+    if (exception_.get() == nullptr) {
+      exception_.reset(new platform::EOFException(exp));
+      type_ = kEOF;
+    }
+  }
+
   enum ExceptionType { kNone, kEnforceNotMet, kEOF };
   ExceptionType type_{kNone};
 

paddle/fluid/framework/details/threaded_ssa_graph_executor.cc

@@ -107,11 +107,11 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
     auto cur_ready_vars = ready_vars.PopAll(1, &timeout);
 
     if (timeout) {
-      if (exception_holder_.ExceptionCatched()) {
+      if (exception_holder_.IsCaught()) {
         for (auto &run_op_future : run_op_futures_) {
           run_op_future.wait();
         }
-        exception_holder_.Throw();
+        exception_holder_.ReThrow();
       } else {
         continue;
       }
@@ -220,12 +220,8 @@ void ThreadedSSAGraphExecutor::RunOp(
       running_ops_--;
       ready_var_q->Extend(op->Outputs());
       VLOG(10) << op << " " << op->Name() << "Signal posted";
-    } catch (platform::EOFException ex) {
-      exception_holder_.Catch(ex);
-    } catch (platform::EnforceNotMet ex) {
-      exception_holder_.Catch(ex);
     } catch (...) {
-      LOG(FATAL) << "Unknown exception catched";
+      exception_holder_.Catch(std::current_exception());
     }
   };
   if (pool_) {

paddle/fluid/framework/ir/CMakeLists.txt

@@ -3,7 +3,10 @@ cc_library(graph SRCS graph.cc DEPS node)
 cc_library(graph_helper SRCS graph_helper.cc DEPS graph)
 cc_library(pass SRCS pass.cc DEPS graph node graph_helper)
 cc_library(graph_viz_pass SRCS graph_viz_pass.cc DEPS graph pass graph_helper)
+cc_library(graph_traits SRCS graph_traits.cc DEPS graph)
+cc_library(graph_pattern_detecter SRCS graph_pattern_detecter.cc DEPS graph graph_helper graph_traits)
 
 cc_test(pass_test SRCS pass_test.cc DEPS graph pass graph_helper)
 cc_test(graph_test SRCS graph_test.cc DEPS graph graph_helper op_registry)
 cc_test(graph_helper_test SRCS graph_helper_test.cc DEPS graph graph_helper op_registry)
+cc_test(test_graph_pattern_detecter SRCS graph_pattern_detecter_tester.cc DEPS graph_pattern_detecter)

paddle/fluid/framework/ir/graph.h

@@ -28,6 +28,38 @@ namespace paddle {
 namespace framework {
 namespace ir {
 
+/*
+ * The graph is a Directed Acyclic Single Static Assignment Graph.
+ *
+ * In more detail, the following properties must hold:
+ *
+ *   The graph shouldn't contain cycle. Each node is a black-box to the graph
+ *   so the node itself could be a loop operator.
+ *
+ *   Each Variable-type node has only one input (thus single static assignment).
+ *
+ *   The output/input of operator is variable and the output/input of variable
+ *   is operator.
+ *
+ * The following data harzards in Program are addressed in the Graph:
+ *
+ *   Write-After-Read
+ *     a = op1(x)
+ *     x = op2(b)
+ *     A control-dependency connection is created bettwen op1 and op2 such that
+ *     op1->op2, so as to ensure correct order.
+ *
+ *   Write-After-Write
+ *     x = op1(a)
+ *     x = op2(b)
+ *     A control-dependency connection is created between op1 and op2 such that
+ *     op1->op2, so as to ensure correct order.
+ *
+ * Other properties currently hold, but is not enforced yet:
+ *
+ *   Variable-type node (not control dep) with the same variable name share
+ *   the same underlying VarDesc.
+ */
 class Graph {
  public:
   explicit Graph(const ProgramDesc &program);