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

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/inference/analysis/analyzer.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_tester.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/argument.h

@@ -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/ir_pass_manager.cc

@@ -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_passes/tensorrt_subgraph_pass.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.h

@@ -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/passes/CMakeLists.txt

@@ -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/inference_op_replace_pass.cc

@@ -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();

paddle/fluid/framework/ir/infer_clean_graph_pass.cc → paddle/fluid/inference/analysis/passes/ir_graph_clean_pass.cc

@@ -12,56 +12,36 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include "paddle/fluid/inference/analysis/passes/ir_graph_clean_pass.h"
 #include <algorithm>
-#include "paddle/fluid/framework/ir/fuse_pass_base.h"
 #include "paddle/fluid/framework/ir/graph.h"
 #include "paddle/fluid/framework/ir/graph_pattern_detector.h"
+#include "paddle/fluid/framework/ir/node.h"
 
 namespace paddle {
-namespace framework {
-namespace ir {
-
-class InferCleanGraphPass : public FusePassBase {
- public:
-  virtual ~InferCleanGraphPass() {}
-
- protected:
-  void ApplyImpl(ir::Graph* graph) const {
-    FusePassBase::Init("original_graph", graph);
-    PADDLE_ENFORCE(graph);
-
-    auto is_valid_node = [](Node* x) {
-      return x && IsControlDepVar(*x) && x->IsVar() && !x->Var();
-    };
-
-    std::unordered_set<const Node*> invalid_nodes;
-    int valid_op = 0;
-    for (auto* node : graph->Nodes()) {
-      PADDLE_ENFORCE_NOT_NULL(node);
-      if (is_valid_node(node)) {
-        invalid_nodes.insert(node);
-      } else if (node->IsOp()) {
-        // Collect all the operators to help tracking number of operators.
-        ++valid_op;
-      }
+namespace inference {
+namespace analysis {
+
+void IrInferCleanGraphPass::RunImpl(Argument* argument) {
+  auto& graph = argument->main_graph();
+  auto is_valid_node = [](framework::ir::Node* x) {
+    return x && IsControlDepVar(*x) && x->IsVar() && !x->Var();
+  };
+
+  std::unordered_set<const framework::ir::Node*> invalid_nodes;
+  int valid_op = 0;
+  for (auto* node : graph.Nodes()) {
+    PADDLE_ENFORCE_NOT_NULL(node);
+    if (is_valid_node(node)) {
+      invalid_nodes.insert(node);
+    } else if (node->IsOp()) {
+      ++valid_op;
     }
-
-    GraphSafeRemoveNodes(graph, invalid_nodes);
-
-    AddStatis(valid_op);
   }
 
-  void CleanEdges(std::vector<Node*>* nodes,
-                  const std::unordered_set<Node*>& to_remove) const {
-    auto it = std::remove_if(nodes->begin(), nodes->end(),
-                             [&](Node* x) { return to_remove.count(x); });
-    nodes->erase(it, nodes->end());
-  }
-};
+  GraphSafeRemoveNodes(&graph, invalid_nodes);
+}
 
-}  // namespace ir
-}  // namespace framework
+}  // namespace analysis
+}  // namespace inference
 }  // namespace paddle
-
-REGISTER_PASS(infer_clean_graph_pass,
-              paddle::framework::ir::InferCleanGraphPass);

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

+// 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 <string>
+#include <unordered_set>
+#include "paddle/fluid/inference/analysis/analysis_pass.h"
+
+namespace paddle {
+namespace inference {
+namespace analysis {
+
+class IrInferCleanGraphPass : public AnalysisPass {
+ public:
+  void RunImpl(Argument *argument) override;
+
+  std::string repr() const override { return "ir_graph_clean_pass"; }
+};
+
+}  // namespace analysis
+}  // namespace inference
+}  // namespace paddle

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

@@ -109,10 +109,16 @@ int DataTypeToSpace(framework::proto::VarType_Type type) {
 void MemoryOptimizePass::CollectVarMemorySize(
     space_table_t* space_table) const {
   const int fake_batch_size = 1;
+
   auto valid_var = [&](framework::ir::Node* node) -> bool {
-    std::set<std::string> invalid_op = {"while", "conditional_block",
+    std::set<std::string> invalid_op = {"while",
+                                        "conditional_block",
                                         "tensorrt_engine",
-                                        "conditional_block_infer"};
+                                        "conditional_block_infer",
+                                        "merge_lod_tensor_infer",
+                                        "merge_lod_tensor",
+                                        "equal",
+                                        "lod_reset"};
     for (auto* tmp : node->inputs) {
       CHECK(tmp->IsOp());
       std::string op_type = tmp->Op()->Type();

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

@@ -75,6 +75,7 @@ class MemoryOptimizePass : public AnalysisPass {
       int sort_kind) const;
 
   void CollectVarMemorySize(space_table_t *space_table) const;
+  void CollectVarMemorySize0(space_table_t *space_table) const;
 
   void CollectVarMemorySize(
       const std::unordered_map<std::string, size_t> &batch_var_ave_dim,

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

@@ -17,6 +17,7 @@
 #include "paddle/fluid/inference/analysis/passes/inference_op_replace_pass.h"
 #include "paddle/fluid/inference/analysis/passes/ir_analysis_pass.h"
 #include "paddle/fluid/inference/analysis/passes/ir_graph_build_pass.h"
+#include "paddle/fluid/inference/analysis/passes/ir_graph_clean_pass.h"
 #include "paddle/fluid/inference/analysis/passes/ir_graph_to_program_pass.h"
 #include "paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.h"
 #include "paddle/fluid/inference/analysis/passes/memory_optimize_pass.h"
@@ -32,6 +33,8 @@ PassRegistry::PassRegistry() {
                   std::unique_ptr<AnalysisPass>(new IrAnalysisPass));
   passes_.emplace("ir_graph_build_pass",
                   std::unique_ptr<AnalysisPass>(new IrGraphBuildPass));
+  passes_.emplace("ir_graph_clean_pass",
+                  std::unique_ptr<AnalysisPass>(new IrInferCleanGraphPass));
   passes_.emplace("memory_optimize_pass",
                   std::unique_ptr<AnalysisPass>(new MemoryOptimizePass));
   passes_.emplace(

paddle/fluid/inference/api/analysis_predictor.cc

@@ -135,7 +135,6 @@ bool AnalysisPredictor::PrepareProgram(
     const std::shared_ptr<framework::ProgramDesc> &program) {
   if (!program) {
     if (!LoadProgramDesc()) return false;
-
     // If not cloned, the parameters should be loaded.
     // If config_.ir_optim() is True, parameters is loaded in
     // OptimizeInferenceProgram(), but other persistable variables
@@ -145,17 +144,10 @@ bool AnalysisPredictor::PrepareProgram(
     // So in both case, create persistable variables at first.
     executor_->CreateVariables(*inference_program_, 0, true, sub_scope_);
 
-    // Optimize the program, and load parameters and modify them in the
-    // scope_.
-    // This will change the scope_ address.
-    if (config_.ir_optim()) {
-      status_ir_optim_enabled_ = true;
-      OptimizeInferenceProgram();
-    } else {
-      // Load parameters
-      LOG(INFO) << "load parameters ";
-      LoadParameters();
-    }
+    // if enable_ir_optim_ is false,
+    // the analysis pass(op fuse, graph analysis, trt subgraph, mkldnn etc) will
+    // not be executed.
+    OptimizeInferenceProgram();
   } else {
     // If the program is passed from external, no need to optimize it, this
     // logic is used in the clone scenario.
@@ -363,6 +355,7 @@ bool AnalysisPredictor::GetFetch(std::vector<PaddleTensor> *outputs,
 void AnalysisPredictor::PrepareArgument() {
   argument_.SetUseGPU(config_.use_gpu());
   argument_.SetGPUDeviceId(config_.gpu_device_id());
+  argument_.SetEnableAnalysisOptim(config_.enable_ir_optim_);
   argument_.SetEnableMemoryOptim(config_.enable_memory_optim());
   argument_.SetStaticMemoryOptim(config_.static_memory_optim_);
   argument_.SetStaticMemoryOptimForceUpdate(
@@ -434,8 +427,6 @@ void AnalysisPredictor::PrepareArgument() {
 
 // NOTE All the members in AnalysisConfig should be copied to Argument.
 void AnalysisPredictor::OptimizeInferenceProgram() {
-  status_program_optimized_ = true;
-
   PrepareArgument();
   Analyzer().Run(&argument_);
 

paddle/fluid/inference/api/analysis_predictor.h

@@ -175,10 +175,8 @@ class AnalysisPredictor : public PaddlePredictor {
 
  private:
   // Some status here that help to determine the status inside the predictor.
-  bool status_program_optimized_{false};
   bool status_is_cloned_{false};
   bool status_use_gpu_{false};
-  bool status_ir_optim_enabled_{false};
 };
 
 }  // namespace paddle

paddle/fluid/inference/api/analysis_predictor_tester.cc

@@ -44,7 +44,6 @@ TEST(AnalysisPredictor, analysis_off) {
   ASSERT_EQ(predictor->scope_->parent(), nullptr);
   ASSERT_EQ(predictor->sub_scope_->parent(), predictor->scope_.get());
   // ir is turned off, so program shouldn't be optimized.
-  ASSERT_FALSE(predictor->status_program_optimized_);
   LOG(INFO) << "scope parameters " << predictor->scope_->LocalVarNames().size();
 
   // 2. Dummy Input Data
@@ -76,8 +75,6 @@ TEST(AnalysisPredictor, analysis_on) {
   ASSERT_TRUE(predictor->sub_scope_);
   ASSERT_EQ(predictor->scope_->parent(), nullptr);
   ASSERT_EQ(predictor->sub_scope_->parent(), predictor->scope_.get());
-  // ir is turned on, so program should be optimized.
-  ASSERT_TRUE(predictor->status_program_optimized_);
   // 2. Dummy Input Data
   int64_t data[4] = {1, 2, 3, 4};
   PaddleTensor tensor;

paddle/fluid/inference/api/mkldnn_quantizer.cc

@@ -367,13 +367,14 @@ void AnalysisPredictor::MkldnnQuantizer::PrepareArgument() const {
 
   auto* builder = predictor_.config_.pass_builder();
   builder->SetPasses({
-      "infer_clean_graph_pass", "cpu_quantize_pass", "cpu_quantize_squash_pass",
+      "cpu_quantize_pass", "cpu_quantize_squash_pass",
   });
   if (predictor_.config_.ir_debug_) builder->TurnOnDebug();
   auto passes = builder->AllPasses();
   predictor_.argument_.SetIrAnalysisPasses(passes);
   predictor_.argument_.SetAnalysisPasses(
-      {"ir_analysis_pass", "memory_optimize_pass", "ir_graph_to_program_pass"});
+      {"ir_graph_clean_pass", "ir_analysis_pass", "memory_optimize_pass",
+       "ir_graph_to_program_pass"});
   predictor_.argument_.SetQuantVarScales(scales_);
 }
 

paddle/fluid/inference/api/paddle_analysis_config.h

@@ -46,6 +46,7 @@ struct AnalysisConfig {
   enum class Precision {
     kFloat32 = 0,
     kInt8,
+    kHalf,
   };
 
   /** Set model with a directory.

paddle/fluid/inference/api/paddle_pass_builder.cc

@@ -71,8 +71,7 @@ void PaddlePassBuilder::AppendAnalysisPass(const std::string &pass) {
 void PaddlePassBuilder::ClearPasses() { passes_.clear(); }
 
 const std::vector<std::string> kTRTSubgraphPasses({
-  "infer_clean_graph_pass",                        //
-      "conv_affine_channel_fuse_pass",             //
+  "conv_affine_channel_fuse_pass",                 //
       "conv_eltwiseadd_affine_channel_fuse_pass",  //
       "quant_conv2d_dequant_fuse_pass",            //
       "delete_quant_dequant_op_pass",              //
@@ -90,7 +89,6 @@ const std::vector<std::string> kTRTSubgraphPasses({
 
 // The following passes works for Anakin sub-graph engine.
 const std::vector<std::string> kAnakinSubgraphPasses({
-    "infer_clean_graph_pass",                       //
     "quant_conv2d_dequant_fuse_pass",               //
     "simplify_anakin_priorbox_detection_out_pass",  //
     "fillconstant_elementwisemul_fuse",             //
@@ -104,9 +102,8 @@ const std::vector<std::string> kAnakinSubgraphPasses({
 
 GpuPassStrategy::GpuPassStrategy() : PassStrategy({}) {
   passes_.assign({
-    "infer_clean_graph_pass",  //
-        //   "identity_scale_op_clean_pass",              //
-        "conv_affine_channel_fuse_pass",             //
+    //   "identity_scale_op_clean_pass",              //
+    "conv_affine_channel_fuse_pass",                 //
         "conv_eltwiseadd_affine_channel_fuse_pass",  //
         "conv_bn_fuse_pass",                         //
         "conv_eltwiseadd_bn_fuse_pass",              //
@@ -140,8 +137,7 @@ void GpuPassStrategy::EnableNgraph() {
 CpuPassStrategy::CpuPassStrategy() : PassStrategy({}) {
   // NOTE the large fusions should be located in the front, so that they will
   // not be damaged by smaller ones.
-  passes_.assign({"infer_clean_graph_pass",         //
-                  "attention_lstm_fuse_pass",       //
+  passes_.assign({"attention_lstm_fuse_pass",       //
                   "seqconv_eltadd_relu_fuse_pass",  //
                   // "seqpool_concat_fuse_pass",    //
                   "seqpool_cvm_concat_fuse_pass",  //

paddle/fluid/inference/api/paddle_pass_builder.h

@@ -72,7 +72,7 @@ class PaddlePassBuilder {
 
  protected:
   std::vector<std::string> analysis_passes_{
-      {"ir_graph_build_pass", "ir_analysis_pass",
+      {"ir_graph_build_pass", "ir_graph_clean_pass", "ir_analysis_pass",
        "ir_params_sync_among_devices_pass", "adjust_cudnn_workspace_size_pass",
        "inference_op_replace_pass"}};
   std::vector<std::string> passes_;

paddle/fluid/inference/tensorrt/convert/test_op_converter.cc

@@ -14,7 +14,7 @@ limitations under the License. */
 
 #include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
 
-#include <gtest/gtest.h>
+#include <gtest/gtest.h>  // NOLINT
 #include "paddle/fluid/framework/program_desc.h"
 
 namespace paddle {
@@ -27,10 +27,8 @@ TEST(OpConverter, ConvertBlock) {
   auto* conv2d_op = block->AppendOp();
 
   // init trt engine
-  cudaStream_t stream_;
   std::unique_ptr<TensorRTEngine> engine_;
-  PADDLE_ENFORCE_EQ(cudaStreamCreate(&stream_), 0);
-  engine_.reset(new TensorRTEngine(5, 1 << 15, stream_));
+  engine_.reset(new TensorRTEngine(5, 1 << 15));
   engine_->InitNetwork();
 
   engine_->DeclareInput("conv2d-X", nvinfer1::DataType::kFLOAT,

paddle/fluid/inference/tensorrt/convert/ut_helper.h

@@ -80,8 +80,7 @@ class TRTConvertValidation {
         if_add_batch_(if_add_batch),
         max_batch_size_(max_batch_size) {
     PADDLE_ENFORCE_EQ(cudaStreamCreate(&stream_), 0);
-    engine_.reset(
-        new TensorRTEngine(max_batch_size, workspace_size, false, nullptr, 0));
+    engine_.reset(new TensorRTEngine(max_batch_size, workspace_size));
     engine_->InitNetwork();
   }
 

paddle/fluid/inference/tensorrt/engine.cc

@@ -51,7 +51,27 @@ void TensorRTEngine::FreezeNetwork() {
   // build engine.
   infer_builder_->setMaxBatchSize(max_batch_);
   infer_builder_->setMaxWorkspaceSize(max_workspace_);
-  if (enable_int8_) {
+  bool enable_fp16 = (precision_ == AnalysisConfig::Precision::kHalf);
+#if IS_TRT_VERSION_GE(5000)
+  if (enable_fp16) {
+    bool support_fp16 = infer_builder_->platformHasFastFp16();
+    infer_builder_->setFp16Mode(support_fp16);
+    if (!support_fp16) {
+      LOG(INFO) << "You specify FP16 mode, but the hardware do not support "
+                   "FP16 speed up, use FP32 instead.";
+    } else {
+      LOG(INFO) << "Run Paddle-TRT FP16 mode. ";
+    }
+  }
+#else
+  if (enable_fp16)
+    LOG(INFO) << "Using FP16 in Paddle-trt must ensure that the version of TRT "
+                 "is at least 5."
+                 " So, use FP32 to run.";
+#endif
+  bool enable_int8 = (precision_ == AnalysisConfig::Precision::kInt8);
+
+  if (enable_int8) {
     infer_builder_->setInt8Mode(true);
     if (calibrator_) {
       infer_builder_->setInt8Calibrator(calibrator_);

paddle/fluid/inference/tensorrt/engine.h

@@ -22,6 +22,7 @@ limitations under the License. */
 #include <vector>
 #include "paddle/fluid/framework/tensor.h"
 #include "paddle/fluid/framework/tensor_util.h"
+#include "paddle/fluid/inference/api/paddle_analysis_config.h"
 #include "paddle/fluid/inference/engine.h"
 #include "paddle/fluid/inference/tensorrt/helper.h"
 #include "paddle/fluid/inference/tensorrt/plugin/trt_plugin.h"
@@ -61,12 +62,14 @@ class TensorRTEngine {
     nvinfer1::Weights w_;
   };
 
-  TensorRTEngine(int max_batch, int max_workspace, bool enable_int8 = false,
-                 TRTInt8Calibrator* calibrator = nullptr, int device_id = 0,
-                 nvinfer1::ILogger& logger = NaiveLogger::Global())
+  TensorRTEngine(
+      int max_batch, int max_workspace,
+      AnalysisConfig::Precision precision = AnalysisConfig::Precision::kFloat32,
+      TRTInt8Calibrator* calibrator = nullptr, int device_id = 0,
+      nvinfer1::ILogger& logger = NaiveLogger::Global())
       : max_batch_(max_batch),
         max_workspace_(max_workspace),
-        enable_int8_(enable_int8),
+        precision_(precision),
         calibrator_(calibrator),
         device_id_(device_id),
         logger_(logger) {}
@@ -168,7 +171,7 @@ class TensorRTEngine {
   // the max memory size the engine uses
   int max_workspace_;
 
-  bool enable_int8_;
+  AnalysisConfig::Precision precision_;
   TRTInt8Calibrator* calibrator_;
   // batch size of the current data, will be updated each Executation.
   int batch_size_{-1};
@@ -231,12 +234,12 @@ class TRTEngineManager {
     return engines_.at(name).get();
   }
 
-  TensorRTEngine* Create(std::string name, int max_batch, int max_workspace,
-                         bool enable_int8 = false,
-                         TRTInt8Calibrator* calibrator = nullptr,
-                         int device_id = 0,
-                         nvinfer1::ILogger& logger = NaiveLogger::Global()) {
-    auto* p = new TensorRTEngine(max_batch, max_workspace, enable_int8,
+  TensorRTEngine* Create(
+      std::string name, int max_batch, int max_workspace,
+      AnalysisConfig::Precision precision = AnalysisConfig::Precision::kFloat32,
+      TRTInt8Calibrator* calibrator = nullptr, int device_id = 0,
+      nvinfer1::ILogger& logger = NaiveLogger::Global()) {
+    auto* p = new TensorRTEngine(max_batch, max_workspace, precision,
                                  calibrator, device_id, logger);
     engines_[name].reset(p);
     return p;

paddle/fluid/operators/merge_lod_tensor_op.cc

@@ -28,9 +28,9 @@ class MergeLoDTensorOp : public framework::OperatorBase {
                    const framework::AttributeMap &attrs)
       : OperatorBase(type, inputs, outputs, attrs) {}
 
- private:
-  void RunImpl(const framework::Scope &scope,
-               const platform::Place &dev_place) const override {
+ protected:
+  void RunBase(const framework::Scope &scope,
+               const platform::Place &dev_place) const {
     // get device context from pool
     platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
     auto &dev_ctx = *pool.Get(dev_place);
@@ -125,6 +125,33 @@ class MergeLoDTensorOp : public framework::OperatorBase {
       out_lod->insert(out_lod->begin(), x.lod()[i]);
     }
   }
+
+ private:
+  void RunImpl(const framework::Scope &scope,
+               const platform::Place &dev_place) const override {
+    RunBase(scope, dev_place);
+  }
+};
+
+class MergeLoDTensorInferOp : public MergeLoDTensorOp {
+ public:
+  MergeLoDTensorInferOp(const std::string &type,
+                        const framework::VariableNameMap &inputs,
+                        const framework::VariableNameMap &outputs,
+                        const framework::AttributeMap &attrs)
+      : MergeLoDTensorOp(type, inputs, outputs, attrs) {}
+
+ private:
+  void RunImpl(const framework::Scope &scope,
+               const platform::Place &dev_place) const override {
+    RunBase(scope, dev_place);
+    framework::Variable *in_true_var = scope.FindVar(Input("InTrue"));
+    framework::Variable *in_false_var = scope.FindVar(Input("InFalse"));
+    in_true_var->Clear();
+    in_false_var->Clear();
+    in_true_var->GetMutable<framework::LoDTensor>();
+    in_false_var->GetMutable<framework::LoDTensor>();
+  }
 };
 
 class MergeLoDTensorOpProtoMaker : public framework::OpProtoAndCheckerMaker {
@@ -196,3 +223,7 @@ namespace ops = paddle::operators;
 REGISTER_OPERATOR(merge_lod_tensor, ops::MergeLoDTensorOp,
                   ops::MergeLoDTensorOpProtoMaker,
                   ops::MergeLoDTensorInferShape, ops::MergeLoDTensorGradMaker);
+REGISTER_OPERATOR(merge_lod_tensor_infer, ops::MergeLoDTensorInferOp,
+                  ops::MergeLoDTensorOpProtoMaker,
+                  ops::MergeLoDTensorInferShape,
+                  paddle::framework::EmptyGradOpMaker);

paddle/fluid/operators/tensorrt/tensorrt_engine_op.h

@@ -48,12 +48,14 @@ class TensorRTEngineOp : public framework::OperatorBase {
   int workspace_size_;
   std::unique_ptr<TRTInt8Calibrator> calibrator_;
   bool enable_int8_;
+  bool enable_fp16_;
   bool use_calib_mode_;
   std::string calibration_data_;
   std::string engine_key_;
   bool calibration_mode_;
   int predictor_id_;
   int device_id_;
+  AnalysisConfig::Precision precision_mode_;
 
  public:
   TensorRTEngineOp(const std::string &type,
@@ -66,6 +68,7 @@ class TensorRTEngineOp : public framework::OperatorBase {
     workspace_size_ = Attr<int>("workspace_size");
     device_id_ = Attr<int>("gpu_id");
     enable_int8_ = Attr<bool>("enable_int8");
+    enable_fp16_ = Attr<bool>("enable_fp16");
     use_calib_mode_ = Attr<bool>("use_calib_mode");
     calibration_data_ = Attr<std::string>("calibration_data");
     engine_key_ = Attr<std::string>("engine_key");
@@ -93,6 +96,13 @@ class TensorRTEngineOp : public framework::OperatorBase {
           inference::Singleton<inference::tensorrt::TRTEngineManager>::Global()
               .Get(engine_key_ + std::to_string(predictor_id_));
     }
+    precision_mode_ = AnalysisConfig::Precision::kFloat32;
+    if (enable_int8_) {
+      precision_mode_ = AnalysisConfig::Precision::kInt8;
+    }
+    if (enable_fp16_) {
+      precision_mode_ = AnalysisConfig::Precision::kHalf;
+    }
   }
 
  protected:
@@ -141,7 +151,7 @@ class TensorRTEngineOp : public framework::OperatorBase {
           calib_buffers, runtime_batch, engine_key_, dev_place));
       calib_res->thr_.reset(new std::thread([&]() {
         calib_res->engine_.reset(new TensorRTEngine(
-            max_batch_size_, workspace_size_, enable_int8_,
+            max_batch_size_, workspace_size_, precision_mode_,
             calib_res->calib_.get(),
             boost::get<platform::CUDAPlace>(dev_place).device));
         VLOG(3) << "start the calib trt engine thread";
@@ -241,7 +251,7 @@ class TensorRTEngineOp : public framework::OperatorBase {
       trt_engine_ =
           inference::Singleton<inference::tensorrt::TRTEngineManager>::Global()
               .Create(engine_key_ + std::to_string(predictor_id_),
-                      max_batch_size_, workspace_size_, enable_int8_,
+                      max_batch_size_, workspace_size_, precision_mode_,
                       calibrator_.get(), device_id_);
       PrepareTRTEngine(scope, trt_engine_);
     }

paddle/fluid/operators/tensorrt/tensorrt_engine_op_test.cc

@@ -105,6 +105,7 @@ TEST(TensorRTEngineOp, manual) {
   engine_op_desc.SetAttr("predictor_id", 1);
   engine_op_desc.SetAttr("calibration_data", std::string(""));
   engine_op_desc.SetAttr("enable_int8", static_cast<bool>(false));
+  engine_op_desc.SetAttr("enable_fp16", static_cast<bool>(false));
   engine_op_desc.SetAttr("use_calib_mode", static_cast<bool>(false));
   engine_op_desc.SetAttr("output_name_mapping",
                          std::vector<std::string>({"z0"}));
@@ -205,6 +206,7 @@ void Execute(int batch_size, int input_dim, int output_dim, int nlayers = 1) {
   engine_op_desc.SetAttr("predictor_id", 1);
   engine_op_desc.SetAttr("calibration_data", std::string(""));
   engine_op_desc.SetAttr("enable_int8", static_cast<bool>(false));
+  engine_op_desc.SetAttr("enable_fp16", static_cast<bool>(false));
   engine_op_desc.SetAttr("use_calib_mode", static_cast<bool>(false));
   engine_op_desc.SetAttr("output_name_mapping",
                          std::vector<std::string>({"z3"}));

paddle/fluid/pybind/inference_api.cc

@@ -199,6 +199,7 @@ void BindAnalysisConfig(py::module *m) {
   py::enum_<AnalysisConfig::Precision>(analysis_config, "Precision")
       .value("Float32", AnalysisConfig::Precision::kFloat32)
       .value("Int8", AnalysisConfig::Precision::kInt8)
+      .value("Half", AnalysisConfig::Precision::kHalf)
       .export_values();
 
   analysis_config.def(py::init<const AnalysisConfig &>())

python/paddle/fluid/tests/unittests/test_split_and_merge_lod_tensor_op.py

@@ -23,6 +23,7 @@ from paddle.fluid.executor import Executor
 from paddle.fluid.backward import append_backward
 from paddle.fluid.layers.control_flow import split_lod_tensor
 from paddle.fluid.layers.control_flow import merge_lod_tensor
+from paddle.fluid.layer_helper import LayerHelper
 
 
 class TestCPULoDTensorArrayOps(unittest.TestCase):
@@ -57,7 +58,7 @@ class TestCPULoDTensorArrayOps(unittest.TestCase):
             expect_false=expect_false,
             expect_out=tensor)
 
-    def test_split_and_merge_lod_tensor_level_0(self):
+    def split_and_merge_lod_tensor_level_0(self, use_merge_lod_infer=False):
         tensor = core.LoDTensor()
         tensor.set(np.arange(10).reshape(10, 1).astype('int32'), self.place())
         tensor.set_recursive_sequence_lengths([[3, 6, 1]])
@@ -87,10 +88,23 @@ class TestCPULoDTensorArrayOps(unittest.TestCase):
             mask=mask,
             expect_true=expect_true,
             expect_false=expect_false,
-            expect_out=tensor)
-
-    def main(self, tensor, mask, expect_true, expect_false, expect_out,
-             level=0):
+            expect_out=tensor,
+            use_merge_lod_infer=use_merge_lod_infer)
+
+    def test_split_and_merge_lod_tensor_1(self):
+        self.split_and_merge_lod_tensor_level_0()
+
+    def test_split_and_merge_lod_tensor_2(self):
+        self.split_and_merge_lod_tensor_level_0(True)
+
+    def main(self,
+             tensor,
+             mask,
+             expect_true,
+             expect_false,
+             expect_out,
+             level=0,
+             use_merge_lod_infer=False):
         place = self.place()
         program = Program()
         with program_guard(program):
@@ -103,11 +117,36 @@ class TestCPULoDTensorArrayOps(unittest.TestCase):
             out_true, out_false = split_lod_tensor(input=x, mask=y, level=level)
             out_true.persistable = True
             out_false.persistable = True
-
-            out = merge_lod_tensor(
-                in_true=out_true, in_false=out_false, mask=y, x=x, level=level)
-
-            out.persistable = True
+            if use_merge_lod_infer:
+                input_dict = {
+                    'X': x,
+                    'Mask': mask,
+                    'InTrue': out_true,
+                    'InFalse': out_false,
+                    'level': level
+                }
+                helper = LayerHelper('merge_lod_tensor_infer')
+                out = helper.create_variable_for_type_inference(
+                    dtype=out_true.dtype)
+                helper.append_op(
+                    type='merge_lod_tensor_infer',
+                    inputs={
+                        'X': x,
+                        'Mask': y,
+                        'InTrue': out_true,
+                        'InFalse': out_false
+                    },
+                    outputs={'Out': out},
+                    attrs={'level': level})
+                out.persistable = True
+            else:
+                out = merge_lod_tensor(
+                    in_true=out_true,
+                    in_false=out_false,
+                    mask=y,
+                    x=x,
+                    level=level)
+                out.persistable = True
 
         exe = Executor(place)
         scope = core.Scope()
@@ -122,9 +161,9 @@ class TestCPULoDTensorArrayOps(unittest.TestCase):
         var_false = scope.find_var(out_false.name).get_tensor()
 
         var_out = scope.find_var(out.name).get_tensor()
-
-        self.check_tensor_same(var_true, expect_true)
-        self.check_tensor_same(var_false, expect_false)
+        if not use_merge_lod_infer:
+            self.check_tensor_same(var_true, expect_true)
+            self.check_tensor_same(var_false, expect_false)
         self.check_tensor_same(var_out, expect_out)
 
     def check_tensor_same(self, actual, expect):