Cherry-pick : anakin subgraph add opt config layout argument #16846

test=release/1.4

paddle/fluid/inference/anakin/engine.cc

@@ -36,13 +36,14 @@ template <typename TargetT, Precision PrecisionType, OpRunType RunType>
 AnakinEngine<TargetT, PrecisionType, RunType>::AnakinEngine(
     bool need_summary, int device, int max_batch_size,
     std::map<std::string, std::vector<int>> max_input_shape,
-    std::vector<std::string> program_inputs)
+    std::vector<std::string> program_inputs, bool auto_config_layout)
     : graph_(new AnakinGraphT<TargetT, PrecisionType>()),
       net_(new AnakinNetT<TargetT, PrecisionType, RunType>(need_summary)) {
   device_ = device;
   max_batch_size_ = max_batch_size;
   max_input_shape_ = max_input_shape;
   program_inputs_ = program_inputs;
+  auto_config_layout_ = auto_config_layout;
 }
 
 template <typename TargetT, Precision PrecisionType, OpRunType RunType>
@@ -57,7 +58,7 @@ void AnakinEngine<TargetT, PrecisionType, RunType>::SetInputShape(
 
 template <typename TargetT, Precision PrecisionType, OpRunType RunType>
 void AnakinEngine<TargetT, PrecisionType, RunType>::InitNet() {
-  net_->init(*graph_);
+  net_->init(*graph_, auto_config_layout_);
 }
 
 template <typename TargetT, Precision PrecisionType, OpRunType RunType>

paddle/fluid/inference/anakin/engine.h

@@ -58,7 +58,8 @@ class AnakinEngine {
   explicit AnakinEngine(
       bool need_summary = false, int device = 0, int max_batch_size = 1,
       std::map<std::string, std::vector<int>> max_input_shape = {},
-      std::vector<std::string> program_inputs = {});
+      std::vector<std::string> program_inputs = {},
+      bool auto_config_layout = false);
   ~AnakinEngine();
   void InitNet();
   void SetInputShape(const std::string &name, std::vector<int> shape);
@@ -120,6 +121,8 @@ class AnakinEngine {
   std::unique_ptr<NetT> net_;
   std::vector<std::string> program_inputs_;
   std::unordered_map<std::string, float> tensor_scales_;
+  // Always be false in gpu mode but true in most cpu cases.
+  bool auto_config_layout_;
 };
 
 template <typename TargetT, ::anakin::Precision PrecisionType>
@@ -138,10 +141,11 @@ class AnakinEngineManager {
   AnakinEngineT *Create(bool need_summary, int device, int max_batch_size,
                         std::map<std::string, std::vector<int>> max_input_shape,
                         std::vector<std::string> program_inputs,
-                        std::string engine_name) {
+                        bool auto_config_layout, std::string engine_name) {
     std::unique_lock<std::mutex> lk(mut_);
     auto *p = new AnakinEngine<TargetT, PrecisionType>(
-        need_summary, device, max_batch_size, max_input_shape, program_inputs);
+        need_summary, device, max_batch_size, max_input_shape, program_inputs,
+        auto_config_layout);
     engines_[engine_name].reset(p);
     return p;
   }

paddle/fluid/inference/analysis/argument.h

@@ -171,6 +171,7 @@ struct Argument {
   DECL_ARGUMENT_FIELD(anakin_min_subgraph_size, AnakinMinSubgraphSize, int);
   DECL_ARGUMENT_FIELD(anakin_precision_mode, AnakinPrecisionMode,
                       AnalysisConfig::Precision);
+  DECL_ARGUMENT_FIELD(anakin_auto_config_layout, AnakinAutoConfigLayout, bool);
   DECL_ARGUMENT_FIELD(use_anakin, UseAnakin, bool);
   DECL_ARGUMENT_FIELD(anakin_passes_filter, AnakinPassesFilter,
                       std::vector<std::string>);

paddle/fluid/inference/analysis/ir_pass_manager.cc

@@ -128,6 +128,8 @@ void IRPassManager::CreatePasses(Argument *argument,
       pass->Set("enable_int8", new bool(enable_int8));
       pass->Set("anakin_ops_filter",
                 new std::vector<std::string>(argument->anakin_ops_filter()));
+      pass->Set("auto_config_layout",
+                new bool(argument->anakin_auto_config_layout()));
     }
 
     pre_pass = pass_name;

paddle/fluid/inference/analysis/ir_passes/anakin_subgraph_pass.cc

@@ -226,18 +226,20 @@ void AnakinSubgraphPass::CreateAnakinEngine(
   auto max_batch_size = Get<int>("max_batch_size");
   auto max_input_shape =
       Get<std::map<std::string, std::vector<int>>>("max_input_shape");
+  bool auto_config_layout = Get<bool>("auto_config_layout");
   if (use_gpu) {
 #ifdef PADDLE_WITH_CUDA
     inference::Singleton<
         anakin::AnakinEngineManager<::anakin::saber::NV, PrecisionT>>::Global()
         .Create(true, Get<int>("gpu_device_id"), max_batch_size,
-                max_input_shape, program_inputs, engine_key);
+                max_input_shape, program_inputs, false, engine_key);
 #endif
   } else {
     inference::Singleton<
         anakin::AnakinEngineManager<::anakin::saber::X86, PrecisionT>>::Global()
         .Create(true, Get<int>("gpu_device_id"), max_batch_size,
-                max_input_shape, program_inputs, engine_key);
+                max_input_shape, program_inputs, auto_config_layout,
+                engine_key);
   }
 
   auto *scope = param_scope();

paddle/fluid/inference/api/analysis_config.cc

@@ -117,6 +117,7 @@ AnalysisConfig::AnalysisConfig(const AnalysisConfig &other) {
   CP_MEMBER(anakin_max_input_shape_);
   CP_MEMBER(anakin_min_subgraph_size_);
   CP_MEMBER(anakin_precision_mode_);
+  CP_MEMBER(anakin_auto_config_layout_);
   CP_MEMBER(anakin_passes_filter_);
   CP_MEMBER(anakin_ops_filter_);
 
@@ -398,7 +399,7 @@ void AnalysisConfig::SwitchIrDebug(int x) {
 void AnalysisConfig::EnableAnakinEngine(
     int max_batch_size, std::map<std::string, std::vector<int>> max_input_shape,
     int min_subgraph_size, AnalysisConfig::Precision precision_mode,
-    std::vector<std::string> passes_filter,
+    bool auto_config_layout, std::vector<std::string> passes_filter,
     std::vector<std::string> ops_filter) {
   anakin_max_batchsize_ = max_batch_size;
   anakin_max_input_shape_ = max_input_shape;
@@ -407,6 +408,7 @@ void AnalysisConfig::EnableAnakinEngine(
   anakin_ops_filter_ = ops_filter;
   use_anakin_ = true;
   anakin_precision_mode_ = precision_mode;
+  anakin_auto_config_layout_ = auto_config_layout;
   Update();
 }
 }  // namespace paddle

paddle/fluid/inference/api/analysis_predictor.cc

@@ -387,6 +387,7 @@ void AnalysisPredictor::PrepareArgument() {
     argument_.SetAnakinMaxInputShape(config_.anakin_max_input_shape_);
     argument_.SetAnakinMinSubgraphSize(config_.anakin_min_subgraph_size_);
     argument_.SetAnakinPrecisionMode(config_.anakin_precision_mode_);
+    argument_.SetAnakinAutoConfigLayout(config_.anakin_auto_config_layout_);
     argument_.SetAnakinPassesFilter(config_.anakin_passes_filter_);
     argument_.SetAnakinOpsFilter(config_.anakin_ops_filter_);
     LOG(INFO) << "Anakin subgraph engine is enabled";

paddle/fluid/inference/api/paddle_analysis_config.h

@@ -153,6 +153,7 @@ struct AnalysisConfig {
       int max_batch_size = 1,
       std::map<std::string, std::vector<int>> max_input_shape = {},
       int min_subgraph_size = 6, Precision precision = Precision::kFloat32,
+      bool auto_config_layout = false,
       std::vector<std::string> passes_filter = {},
       std::vector<std::string> ops_filter = {});
 
@@ -294,6 +295,7 @@ struct AnalysisConfig {
   int anakin_min_subgraph_size_{6};
   std::map<std::string, std::vector<int>> anakin_max_input_shape_;
   Precision anakin_precision_mode_;
+  bool anakin_auto_config_layout_{false};
   std::vector<std::string> anakin_passes_filter_;
   std::vector<std::string> anakin_ops_filter_;
   std::map<std::string, std::string> engine_opt_info_;

paddle/fluid/inference/api/paddle_pass_builder.cc

@@ -78,8 +78,6 @@ const std::vector<std::string> kAnakinSubgraphPasses({
     "fillconstant_elementwisemul_fuse",             //
     "fc_fuse_pass",                                 //
     "conv_elementwise_add_fuse_pass",               //
-    // "conv_bn_fuse_pass",                            //
-    // "conv_elementwise_add_fuse_pass",               //
     "fc_gru_fuse_pass",                             //
     "anakin_subgraph_pass",                         //
     "fc_gru_fuse_pass",                             //

paddle/fluid/pybind/inference_api.cc

@@ -236,6 +236,7 @@ void BindAnalysisConfig(py::module *m) {
                std::map<std::string, std::vector<int>>(),
            py::arg("min_subgraph_size") = 6,
            py::arg("precision_mode") = AnalysisConfig::Precision::kFloat32,
+           py::arg("auto_config_layout") = false,
            py::arg("passes_filter") = std::vector<std::string>(),
            py::arg("ops_filter") = std::vector<std::string>())
       .def("tensorrt_engine_enabled", &AnalysisConfig::tensorrt_engine_enabled)