CHERRY PICK FROM 18941, 18860, 19213：Fix Mask RCNN bug AND Paddle-TRT fp16 support (#19378)

* CHERRY_PICK 18941, 18860: TRT fp16 support.

test=release/1.5

* CHERRY_PICK 19213: Fix BUG: Mask RCNN inference diff When using AnalysisPredictor.
    1. fix affine channel fuse pass.
    2. fix condition block op.
    3. fix merge lod tensor op bug.
    4. fix memory optim cause by reset lod op.

    test=release/1.5

Author: NHZlX

paddle/fluid/framework/ir/CMakeLists.txt

@@ -52,7 +52,6 @@ pass_library(graph_viz_pass base)
 pass_library(lock_free_optimize_pass base)
 pass_library(fc_fuse_pass inference)
 pass_library(attention_lstm_fuse_pass inference)
-pass_library(infer_clean_graph_pass inference)
 pass_library(fc_lstm_fuse_pass inference)
 pass_library(embedding_fc_lstm_fuse_pass inference)
 pass_library(fc_gru_fuse_pass inference)

paddle/fluid/framework/ir/infer_clean_graph_pass.cc

@@ -31,6 +31,9 @@ void Analyzer::RunAnalysis(Argument *argument) {
                  "analsis_passes is not valid in the argument.");
   for (auto &pass : argument->analysis_passes()) {
     string::PrettyLogH1("--- Running analysis [%s]", pass);
+    if (!argument->enable_analysis_optim() && pass == "ir_analysis_pass")
+      continue;
+
     auto *ptr = PassRegistry::Global().Retreive(pass);
     PADDLE_ENFORCE_NOT_NULL(ptr, "no analysis pass called %s", pass);
     ptr->Run(argument);

paddle/fluid/inference/analysis/analyzer.cc

@@ -30,7 +30,7 @@ using namespace framework;  // NOLINT
 TEST(Analyzer, analysis_without_tensorrt) {
   Argument argument;
   argument.SetModelDir(FLAGS_inference_model_dir);
-  argument.SetIrAnalysisPasses({"infer_clean_graph_pass"});
+  argument.SetEnableAnalysisOptim(false);
   argument.SetUseGPU(false);
   argument.SetAnalysisPasses({"ir_graph_build_pass", "ir_analysis_pass",
                               "ir_params_sync_among_devices_pass"});
@@ -41,10 +41,10 @@ TEST(Analyzer, analysis_without_tensorrt) {
 
 TEST(Analyzer, analysis_with_tensorrt) {
   Argument argument;
+  argument.SetEnableAnalysisOptim(false);
   argument.SetTensorRtMaxBatchSize(3);
   argument.SetTensorRtWorkspaceSize(1 << 20);
   argument.SetModelDir(FLAGS_inference_model_dir);
-  argument.SetIrAnalysisPasses({"infer_clean_graph_pass"});
   argument.SetUseGPU(false);
   argument.SetAnalysisPasses({"ir_graph_build_pass", "ir_analysis_pass",
                               "ir_params_sync_among_devices_pass"});

paddle/fluid/inference/analysis/analyzer_tester.cc

@@ -62,6 +62,9 @@ struct Argument {
   using anakin_max_shape_t = std::map<std::string, std::vector<int>>;
 
   bool Has(const std::string& key) const { return valid_fields_.count(key); }
+  // If we set the model using config.SetModelBuffer,
+  // the model and parameter will occupy additional CPU resources.
+  // Use this interface to release these resources.
   void PartiallyRelease() {
     if (Has("model_program_path")) {
       if (Has("model_from_memory") && model_from_memory()) {
@@ -130,6 +133,7 @@ struct Argument {
   DECL_ARGUMENT_FIELD(model_params_path, ModelParamsPath, std::string);
   DECL_ARGUMENT_FIELD(model_from_memory, ModelFromMemory, bool);
   DECL_ARGUMENT_FIELD(optim_cache_dir, OptimCacheDir, std::string);
+  DECL_ARGUMENT_FIELD(enable_analysis_optim, EnableAnalysisOptim, bool);
 
   // The overall graph to work on.
   DECL_ARGUMENT_UNIQUE_FIELD(main_graph, MainGraph, framework::ir::Graph);

paddle/fluid/inference/analysis/argument.h

@@ -84,13 +84,15 @@ void IRPassManager::CreatePasses(Argument *argument,
       pass->Set("program",
                 new framework::ProgramDesc *(&argument->main_program()));
 
-      bool enable_int8 = argument->tensorrt_precision_mode() ==
-                         AnalysisConfig::Precision::kInt8;
+      auto precision_mode = argument->tensorrt_precision_mode();
+      bool enable_int8 = precision_mode == AnalysisConfig::Precision::kInt8;
 
       pass->Set("predictor_id", new int(argument->predictor_id()));
       bool use_calib_mode = argument->tensorrt_use_calib_mode();
       pass->Set("enable_int8", new bool(enable_int8));
       pass->Set("use_calib_mode", new bool(use_calib_mode));
+      pass->Set("precision_mode",
+                new AnalysisConfig::Precision(precision_mode));
 
       bool use_static_engine = argument->tensorrt_use_static_engine();
       bool model_from_memory = argument->model_from_memory();

paddle/fluid/inference/analysis/ir_pass_manager.cc

@@ -149,6 +149,9 @@ void TensorRtSubgraphPass::CreateTensorRTOp(
       graph_var_map[node->Name()] = node;
     }
   }
+  auto precision_mode = Get<AnalysisConfig::Precision>("precision_mode");
+  bool enable_fp16 = false;
+  if (precision_mode == AnalysisConfig::Precision::kHalf) enable_fp16 = true;
   auto enable_int8 = Get<bool>("enable_int8");
   auto use_calib_mode = Get<bool>("use_calib_mode");
   auto &subgraph_nodes = *Agent(node).subgraph();
@@ -216,6 +219,7 @@ void TensorRtSubgraphPass::CreateTensorRTOp(
   SetAttr(op_desc->Proto(), "calibration_data", calibration_data);
 
   SetAttr(op_desc->Proto(), "enable_int8", enable_int8);
+  SetAttr(op_desc->Proto(), "enable_fp16", enable_fp16);
   SetAttr(op_desc->Proto(), "use_calib_mode", use_calib_mode);
   SetAttr(op_desc->Proto(), "engine_key", engine_key);
   SetAttr(op_desc->Proto(), "predictor_id", predictor_id);
@@ -244,7 +248,7 @@ void TensorRtSubgraphPass::CreateTensorRTOp(
       inference::Singleton<inference::tensorrt::TRTEngineManager>::Global()
           .Create(engine_key + std::to_string(predictor_id),
                   Get<int>("max_batch_size"), Get<int>("workspace_size"),
-                  enable_int8, calibrator.get(), Get<int>("gpu_device_id"));
+                  precision_mode, calibrator.get(), Get<int>("gpu_device_id"));
 
   bool need_serialize = (use_static_engine && !load_from_memory);
   if (need_serialize) {

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

@@ -21,6 +21,7 @@
 #include "paddle/fluid/framework/ir/fuse_pass_base.h"
 #include "paddle/fluid/framework/ir/pass.h"
 #include "paddle/fluid/inference/analysis/ir_passes/subgraph_util.h"
+#include "paddle/fluid/inference/api/paddle_analysis_config.h"
 
 namespace paddle {
 namespace inference {

paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.h

@@ -5,6 +5,7 @@ cc_library(ir_params_sync_among_devices_pass SRCS ir_params_sync_among_devices_p
 cc_library(ir_graph_to_program_pass SRCS ir_graph_to_program_pass.cc DEPS analysis_pass graph_to_program_pass)
 cc_library(adjust_cudnn_workspace_size_pass SRCS adjust_cudnn_workspace_size_pass.cc DEPS analysis_pass graph_to_program_pass)
 cc_library(inference_op_replace_pass SRCS inference_op_replace_pass.cc DEPS analysis_pass graph_to_program_pass)
+cc_library(ir_graph_clean_pass SRCS ir_graph_clean_pass.cc DEPS analysis_pass)
 
 cc_library(analysis_passes SRCS passes.cc DEPS
   ir_graph_build_pass
@@ -14,6 +15,7 @@ cc_library(analysis_passes SRCS passes.cc DEPS
   memory_optim_pass
   inference_op_replace_pass
   ir_graph_to_program_pass
+  ir_graph_clean_pass
 )
 
 set(analysis_deps ${analysis_deps}

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

@@ -20,9 +20,9 @@ namespace inference {
 namespace analysis {
 
 void InferenceOpReplacePass::RunImpl(Argument* argument) {
-  if (!argument->use_gpu()) return;
   std::unordered_map<std::string, std::string> replaced_map{
       {"conditional_block", "conditional_block_infer"},
+      {"merge_lod_tensor", "merge_lod_tensor_infer"},
   };
 
   auto& graph = argument->main_graph();