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

Author: NHZlX

paddle/fluid/framework/details/exception_holder.h

@@ -0,0 +1,83 @@
+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include "paddle/fluid/platform/enforce.h"
+
+namespace paddle {
+namespace framework {
+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;
+    }
+  }
+
+  bool ExceptionCatched() const {
+    std::lock_guard<std::mutex> lock(mu_);
+    return exception_.get() != nullptr;
+  }
+
+  void Throw() {
+    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;
+  }
+
+  void Clear() {
+    std::lock_guard<std::mutex> lock(mu_);
+    exception_.reset();
+    type_ = kNone;
+  }
+
+ private:
+  enum ExceptionType { kNone, kEnforceNotMet, kEOF };
+  ExceptionType type_{kNone};
+
+  std::unique_ptr<std::exception> exception_;
+  mutable std::mutex mu_;
+};
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h

@@ -41,7 +41,9 @@ class ScopeBufferedSSAGraphExecutor : public SSAGraphExecutor {
       std::vector<VariableInfo> var_infos, std::vector<platform::Place> places,
       std::unique_ptr<SSAGraphExecutor>&& underlying_executor);
 
-  const ir::Graph& Graph() const { return underlying_executor_->Graph(); }
+  const ir::Graph& Graph() const override {
+    return underlying_executor_->Graph();
+  }
 
   FeedFetchList Run(const std::vector<std::string>& fetch_tensors) override;
 

paddle/fluid/framework/details/threaded_ssa_graph_executor.cc

@@ -83,7 +83,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
 
   // Clean run context
   run_op_futures_.clear();
-  exception_.reset();
+  exception_holder_.Clear();
 
   // Step 3. Execution
   while (!pending_vars.empty()) {
@@ -103,23 +103,11 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
     auto cur_ready_vars = ready_vars.PopAll(1, &timeout);
 
     if (timeout) {
-      std::unique_lock<std::mutex> l(exception_mu_);
-      if (exception_) {
-        l.unlock();
+      if (exception_holder_.ExceptionCatched()) {
         for (auto &run_op_future : run_op_futures_) {
           run_op_future.wait();
         }
-        l.lock();
-        std::exception *exp = exception_.get();
-        if (dynamic_cast<platform::EOFException *>(exp)) {
-          auto e = *static_cast<platform::EOFException *>(exp);
-          throw e;
-        } else if (dynamic_cast<platform::EnforceNotMet *>(exp)) {
-          auto e = *static_cast<platform::EnforceNotMet *>(exp);
-          throw e;
-        } else {
-          LOG(FATAL) << "Unknown exception.";
-        }
+        exception_holder_.Throw();
       } else {
         continue;
       }
@@ -229,14 +217,9 @@ void ThreadedSSAGraphExecutor::RunOp(
       ready_var_q->Extend(op->Outputs());
       VLOG(10) << op << " " << op->Name() << "Signal posted";
     } catch (platform::EOFException ex) {
-      std::lock_guard<std::mutex> l(exception_mu_);
-      // EOFException will not cover up existing EnforceNotMet.
-      if (exception_.get() == nullptr) {
-        exception_.reset(new platform::EOFException(ex));
-      }
+      exception_holder_.Catch(ex);
     } catch (platform::EnforceNotMet ex) {
-      std::lock_guard<std::mutex> l(exception_mu_);
-      exception_.reset(new platform::EnforceNotMet(ex));
+      exception_holder_.Catch(ex);
     } catch (...) {
       LOG(FATAL) << "Unknown exception catched";
     }

paddle/fluid/framework/details/threaded_ssa_graph_executor.h

@@ -24,6 +24,7 @@
 #include <functional>
 #include "ThreadPool.h"  // ThreadPool in thrird party
 #include "paddle/fluid/framework/blocking_queue.h"
+#include "paddle/fluid/framework/details/exception_holder.h"
 #include "paddle/fluid/framework/details/execution_strategy.h"
 #include "paddle/fluid/framework/details/fetch_op_handle.h"
 #include "paddle/fluid/framework/details/ssa_graph_executor.h"
@@ -42,7 +43,7 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
                            const std::vector<platform::Place> &places,
                            std::unique_ptr<ir::Graph> &&graph);
 
-  const ir::Graph &Graph() const { return *graph_; }
+  const ir::Graph &Graph() const override { return *graph_; }
   // Run a SSAGraph by a thread pool
   // Use topological sort algorithm
   FeedFetchList Run(const std::vector<std::string> &fetch_tensors) override;
@@ -59,8 +60,7 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
   std::vector<Scope *> local_scopes_;
   std::vector<platform::Place> places_;
   platform::DeviceContextPool fetch_ctxs_;
-  std::mutex exception_mu_;
-  std::unique_ptr<std::exception> exception_;
+  ExceptionHolder exception_holder_;
   std::atomic<int> running_ops_;
 
   void InsertPendingOp(std::unordered_map<OpHandleBase *, size_t> *pending_ops,

paddle/fluid/inference/analysis/CMakeLists.txt

@@ -6,9 +6,11 @@ cc_library(analysis SRCS pass_manager.cc dot.cc node.cc data_flow_graph.cc graph
   tensorrt_subgraph_node_mark_pass.cc
   analyzer.cc
   helper.cc
+        model_store_pass.cc
   DEPS framework_proto proto_desc)
 cc_test(test_node SRCS node_tester.cc DEPS analysis)
 cc_test(test_dot SRCS dot_tester.cc DEPS analysis)
+cc_binary(inference_analyzer SRCS analyzer_main.cc DEPS analysis)
 
 set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests)
 
@@ -40,3 +42,4 @@ inference_analysis_test(test_tensorrt_subgraph_pass SRCS tensorrt_subgraph_pass_
 inference_analysis_test(test_pass_manager SRCS pass_manager_tester.cc)
 inference_analysis_test(test_tensorrt_subgraph_node_mark_pass SRCS tensorrt_subgraph_node_mark_pass_tester.cc)
 inference_analysis_test(test_analyzer SRCS analyzer_tester.cc)
+inference_analysis_test(test_model_store_pass SRCS model_store_pass_tester.cc)

paddle/fluid/inference/analysis/analyzer.cc

@@ -17,6 +17,7 @@
 #include "paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h"
 #include "paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h"
 #include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h"
+#include "paddle/fluid/inference/analysis/model_store_pass.h"
 #include "paddle/fluid/inference/analysis/pass_manager.h"
 #include "paddle/fluid/inference/analysis/tensorrt_subgraph_node_mark_pass.h"
 #include "paddle/fluid/inference/analysis/tensorrt_subgraph_pass.h"
@@ -29,6 +30,9 @@ DEFINE_bool(inference_analysis_enable_tensorrt_subgraph_engine, false,
 DEFINE_string(inference_analysis_graphviz_log_root, "./",
               "Graphviz debuger for data flow graphs.");
 
+DEFINE_string(inference_analysis_output_storage_path, "",
+              "optimized model output path");
+
 namespace inference {
 namespace analysis {
 
@@ -47,6 +51,9 @@ class DfgPassManagerImpl final : public DfgPassManager {
       AddPass("tensorrt-subgraph", new TensorRTSubGraphPass(trt_teller));
     }
     AddPass("data-flow-graph-to-fluid", new DataFlowGraphToFluidPass);
+    if (!FLAGS_inference_analysis_output_storage_path.empty()) {
+      AddPass("model-store-pass", new ModelStorePass);
+    }
   }
 
   std::string repr() const override { return "dfg-pass-manager"; }

paddle/fluid/inference/analysis/analyzer.h

@@ -16,28 +16,23 @@ limitations under the License. */
 
 /*
  * This file contains Analyzer, an class that exposed as a library that analyze
- * and optimize
- * Fluid ProgramDesc for inference. Similar to LLVM, it has multiple flags to
- * control whether
- * an process is applied on the program.
+ * and optimize Fluid ProgramDesc for inference. Similar to LLVM, it has
+ * multiple flags to
+ * control whether an process is applied on the program.
  *
  * The processes are called Passes in analysis, the Passes are placed in a
- * pipeline, the first
- * Pass is the FluidToDataFlowGraphPass which transforms a Fluid ProgramDesc to
- * a data flow
- * graph, the last Pass is DataFlowGraphToFluidPass which transforms a data flow
- * graph to a
- * Fluid ProgramDesc. The passes in the middle of the pipeline can be any Passes
- * which take a
- * node or data flow graph as input.
+ * pipeline, the first Pass is the FluidToDataFlowGraphPass which transforms a
+ * Fluid ProgramDesc to
+ * a data flow graph, the last Pass is DataFlowGraphToFluidPass which transforms
+ * a data flow graph to a Fluid ProgramDesc. The passes in the middle of the
+ * pipeline can be any Passes
+ * which take a node or data flow graph as input.
  *
  * The Analyzer can be used in two methods, the first is a executable file which
- * can be used to
- * pre-process the inference model and can be controlled by passing difference
- * command flags;
+ * can be used to pre-process the inference model and can be controlled by
+ * passing difference command flags;
  * the other way is to compose inside the inference API as a runtime pre-process
- * phase in the
- * inference service.
+ * phase in the inference service.
  */
 
 #include <gflags/gflags.h>
@@ -50,6 +45,7 @@ namespace paddle {
 // flag if not available.
 DECLARE_bool(inference_analysis_enable_tensorrt_subgraph_engine);
 DECLARE_string(inference_analysis_graphviz_log_root);
+DECLARE_string(inference_analysis_output_storage_path);
 
 namespace inference {
 namespace analysis {

paddle/fluid/inference/analysis/analyzer_main.cc

@@ -0,0 +1,33 @@
+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+/*
+ * This file implements analysizer -- an executation help to analyze and
+ * optimize trained model.
+ */
+#include "paddle/fluid/inference/analysis/analyzer.h"
+#include <gflags/gflags.h>
+#include <glog/logging.h>
+
+int main(int argc, char** argv) {
+  google::ParseCommandLineFlags(&argc, &argv, true);
+  using paddle::inference::analysis::Analyzer;
+  using paddle::inference::analysis::Argument;
+
+  Argument argument;
+  Analyzer analyzer;
+  analyzer.Run(&argument);
+
+  return 0;
+}

paddle/fluid/inference/analysis/analyzer_tester.cc

@@ -20,14 +20,18 @@ namespace paddle {
 namespace inference {
 namespace analysis {
 
-TEST_F(DFG_Tester, analysis_without_tensorrt) {
+TEST(Analyzer, analysis_without_tensorrt) {
   FLAGS_inference_analysis_enable_tensorrt_subgraph_engine = false;
+  Argument argument;
+  argument.fluid_model_dir.reset(new std::string(FLAGS_inference_model_dir));
   Analyzer analyser;
   analyser.Run(&argument);
 }
 
-TEST_F(DFG_Tester, analysis_with_tensorrt) {
+TEST(Analyzer, analysis_with_tensorrt) {
   FLAGS_inference_analysis_enable_tensorrt_subgraph_engine = true;
+  Argument argument;
+  argument.fluid_model_dir.reset(new std::string(FLAGS_inference_model_dir));
   Analyzer analyser;
   analyser.Run(&argument);
 }

paddle/fluid/inference/analysis/argument.h

@@ -36,6 +36,16 @@ namespace analysis {
  * All the fields should be registered here for clearness.
  */
 struct Argument {
+  Argument() = default;
+  explicit Argument(const std::string& fluid_model_dir)
+      : fluid_model_dir(new std::string(fluid_model_dir)) {}
+  // The directory of the trained model.
+  std::unique_ptr<std::string> fluid_model_dir;
+  // The path of `__model__` and `param`, this is used when the file name of
+  // model and param is changed.
+  std::unique_ptr<std::string> fluid_model_program_path;
+  std::unique_ptr<std::string> fluid_model_param_path;
+
   // The graph that process by the Passes or PassManagers.
   std::unique_ptr<DataFlowGraph> main_dfg;
 
@@ -44,6 +54,9 @@ struct Argument {
 
   // The processed program desc.
   std::unique_ptr<framework::proto::ProgramDesc> transformed_program_desc;
+
+  // The output storage path of ModelStorePass.
+  std::unique_ptr<std::string> model_output_store_path;
 };
 
 #define UNLIKELY(condition) __builtin_expect(static_cast<bool>(condition), 0)