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

test=develop

Author: hjchen2

.gitignore

@@ -1,3 +1,4 @@
+python/paddle/fluid/tests/unittests/reader_reset_test.recordio
 paddle/operators/check_t.save
 paddle/operators/check_tensor.ls
 paddle/operators/tensor.save

AUTHORS.md

@@ -42,6 +42,7 @@
 | QiJune | Jun Qi |
 | qingqing01 | Qing-Qing Dang |
 | reyoung | Yang Yu |
+| Sand3r- | Michal Gallus |
 | Superjom | Chun-Wei Yan |
 | tensor-tang | Jian Tang |
 | tianbingsz | Tian-Bing Xu |

cmake/inference_lib.cmake

@@ -166,8 +166,8 @@ copy(framework_lib DEPS ${framework_lib_deps}
 
 set(module "memory")
 copy(memory_lib
-        SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/detail/*.h
-        DSTS ${dst_dir}/${module} ${dst_dir}/${module}/detail
+        SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/detail/*.h ${src_dir}/${module}/allocation/*.h
+        DSTS ${dst_dir}/${module} ${dst_dir}/${module}/detail ${dst_dir}/${module}/allocation
         )
 
 set(inference_deps paddle_fluid_shared paddle_fluid)

paddle/fluid/framework/operator.h

@@ -100,6 +100,7 @@ class OperatorBase {
 
   const std::string& Type() const { return type_; }
 
+  bool HasAttr(const std::string& name) const { return attrs_.count(name); }
   template <typename T>
   inline const T& Attr(const std::string& name) const {
     PADDLE_ENFORCE(attrs_.count(name) != 0, "%s should be in AttributeMap",

paddle/fluid/inference/analysis/CMakeLists.txt

@@ -7,16 +7,17 @@ set(analysis_deps # analysis_deps can be extended accross the project
 add_subdirectory(ir_passes)
 add_subdirectory(passes)
 
-cc_library(ir_pass_manager SRCS ir_pass_manager.cc DEPS graph pass ${INFER_IR_PASSES})
+cc_library(analysis_helper SRCS helper.cc DEPS framework_proto proto_desc graph paddle_fluid_api)
+
+cc_library(ir_pass_manager SRCS ir_pass_manager.cc DEPS graph pass ${INFER_IR_PASSES} analysis_helper)
 
 cc_library(argument SRCS argument.cc DEPS scope proto_desc)
 cc_library(analysis_pass SRCS analysis_pass.cc DEPS proto_desc)
 
 cc_library(analysis SRCS
   analyzer.cc
-  helper.cc
   analysis_pass
-  DEPS ${analysis_deps}
+  DEPS ${analysis_deps} analysis_helper
   )
 
 cc_test(test_dot SRCS dot_tester.cc DEPS analysis)

paddle/fluid/inference/analysis/analyzer_tester.cc

@@ -30,6 +30,7 @@ TEST(Analyzer, analysis_without_tensorrt) {
   Argument argument;
   argument.SetModelDir(FLAGS_inference_model_dir);
   argument.SetIrAnalysisPasses({"infer_clean_graph_pass"});
+  argument.SetUseGPU(false);
 
   Analyzer analyser;
   analyser.Run(&argument);
@@ -41,6 +42,7 @@ TEST(Analyzer, analysis_with_tensorrt) {
   argument.SetTensorRtWorkspaceSize(1 << 20);
   argument.SetModelDir(FLAGS_inference_model_dir);
   argument.SetIrAnalysisPasses({"infer_clean_graph_pass"});
+  argument.SetUseGPU(false);
 
   Analyzer analyser;
   analyser.Run(&argument);

paddle/fluid/inference/analysis/argument.h

@@ -116,6 +116,7 @@ struct Argument {
                       std::vector<std::string>);
 
   DECL_ARGUMENT_FIELD(use_gpu, UseGPU, bool);
+  DECL_ARGUMENT_FIELD(gpu_device_id, GPUDeviceId, int);
   DECL_ARGUMENT_FIELD(use_tensorrt, UseTensorRT, bool);
   DECL_ARGUMENT_FIELD(tensorrt_node_teller, TensorRtNodeTeller,
                       std::function<bool(const framework::ir::Node*)>);

paddle/fluid/inference/analysis/ir_passes/CMakeLists.txt

@@ -4,4 +4,6 @@ set(analysis_deps ${analysis_deps}
         subgraph_detector tensorrt_subgraph_pass
         CACHE INTERNAL "")
 
+set(pass_file ${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/paddle_inference_pass.h)
+file(APPEND ${pass_file} "USE_PASS(tensorrt_subgraph_pass);\n")
 set(INFER_IR_PASSES ${INFER_IR_PASSES} tensorrt_subgraph_pass CACHE INTERNAL "")

paddle/fluid/inference/analysis/passes/ir_graph_build_pass.cc

@@ -30,15 +30,28 @@ void IrGraphBuildPass::RunImpl(Argument *argument) {
   if (!argument->scope_valid()) {
     argument->SetScope(new framework::Scope);
   }
+  PADDLE_ENFORCE(argument->use_gpu_valid());
+
+  // The load program should run on the same device with the inference program,
+  // so that the parameters will on the same device, or they will keep copying
+  // between difference devices.
+  platform::Place place;
+  if (argument->use_gpu()) {
+    PADDLE_ENFORCE(argument->gpu_device_id_valid());
+    place = platform::CUDAPlace(argument->gpu_device_id());
+  } else {
+    place = platform::CPUPlace();
+  }
 
   if (argument->model_dir_valid()) {
-    auto program = LoadModel(argument->model_dir(), argument->scope_ptr());
+    auto program =
+        LoadModel(argument->model_dir(), argument->scope_ptr(), place);
     argument->SetMainProgram(program.release());
   } else if (argument->model_program_path_valid() &&
              argument->model_params_path_valid()) {
     auto program =
         LoadModel(argument->model_program_path(), argument->model_params_path(),
-                  argument->scope_ptr());
+                  argument->scope_ptr(), place);
     argument->SetMainProgram(program.release());
   } else {
     PADDLE_THROW(
@@ -52,16 +65,15 @@ void IrGraphBuildPass::RunImpl(Argument *argument) {
 }
 
 std::unique_ptr<framework::ProgramDesc> IrGraphBuildPass::LoadModel(
-    const std::string &path, framework::Scope *scope) {
-  platform::CPUPlace place;
+    const std::string &path, framework::Scope *scope,
+    const platform::Place &place) {
   framework::Executor exe(place);
   return Load(&exe, scope, path);
 }
 
 std::unique_ptr<framework::ProgramDesc> IrGraphBuildPass::LoadModel(
     const std::string &program_path, const std::string &params_path,
-    framework::Scope *scope) {
-  platform::CPUPlace place;
+    framework::Scope *scope, const platform::Place &place) {
   framework::Executor exe(place);
   return Load(&exe, scope, program_path, params_path);
 }

paddle/fluid/inference/analysis/passes/ir_graph_build_pass.h

@@ -17,6 +17,7 @@
 #include <string>
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/inference/analysis/analysis_pass.h"
+#include "paddle/fluid/platform/place.h"
 
 namespace paddle {
 namespace inference {
@@ -32,11 +33,12 @@ class IrGraphBuildPass : public AnalysisPass {
   std::string repr() const override;
 
  private:
-  std::unique_ptr<framework::ProgramDesc> LoadModel(const std::string &path,
-                                                    framework::Scope *scope);
+  std::unique_ptr<framework::ProgramDesc> LoadModel(
+      const std::string &path, framework::Scope *scope,
+      const platform::Place &place);
   std::unique_ptr<framework::ProgramDesc> LoadModel(
       const std::string &program_path, const std::string &params_path,
-      framework::Scope *scope);
+      framework::Scope *scope, const platform::Place &place);
 
   std::string model_binary_str_;
 };