[cherry-pick][Inference] support mixed precision inference (#49077)

* [Release2.4] Revert python link prs (#48573)

* Revert "Fix mac link python (#48017)"

This reverts commit 3fa7a736.

* Revert "[Cherry-pick] Fix python link error (#47811)"

This reverts commit ff642c68.

* Update config.go

* [Paddle Inference] Add float_to_half_pass to support  inference with mixed precision (#47993)

* [Inference] optimize some code and fix some bug (#48780)

* clean ir_pass_manager and fix map_depthwise_conv_to_conv_pass

* fix unitest timeout

* [Paddle Inference] clean unused code  (#48392)

* fix

* update

* update
Co-authored-by: NChen Weihang <chenweihang@baidu.com>

paddle/fluid/framework/ir/CMakeLists.txt

@@ -148,6 +148,7 @@ pass_library(delete_c_identity_op_pass inference)
 pass_library(preln_residual_bias_fuse_pass inference)
 pass_library(delete_fill_constant_op_pass inference)
 pass_library(constant_folding_pass inference)
+pass_library(auto_mixed_precision_pass inference)
 pass_library(simplify_with_basic_ops_pass base)
 pass_library(fc_elementwise_layernorm_fuse_pass base)
 pass_library(skip_layernorm_fuse_pass base)

paddle/fluid/framework/ir/auto_mixed_precision_pass.h

+// Copyright (c) 2022 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_map>
+#include <unordered_set>
+
+#include "paddle/fluid/framework/ir/fuse_pass_base.h"
+#include "paddle/fluid/framework/ir/node.h"
+#include "paddle/phi/common/backend.h"
+#include "paddle/phi/common/data_type.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+class AutoMixedPrecisionPass : public FusePassBase {
+ public:
+  using VarType = framework::proto::VarType;
+
+ public:
+  AutoMixedPrecisionPass() = default;
+  ~AutoMixedPrecisionPass() = default;
+
+ protected:
+  void ApplyImpl(Graph* graph) const override;
+
+ private:
+  void Init(Graph* graph) const;
+
+  void SetDefaultBlacklist() const;
+
+  void SetOpUniqueType() const;
+
+  void RestoreOpOriginType() const;
+
+  inline std::string GetOpOriginalType(const std::string& op_type) const;
+
+  void GetOpPrecision() const;
+
+  void UpdateOpPrecision() const;
+
+  void InsertCastOp() const;
+
+  void ProcessOpWithDtypeAttr() const;
+
+  bool InputVarsNotConvert(Node* op_node, const std::string& var_name) const;
+
+  bool OutputVarsNotConvert(Node* op_node, const std::string& var_name) const;
+
+  void SetVarPrecision() const;
+
+  void ConvertWeightsData() const;
+
+ private:
+  mutable bool skip_pass_{false};
+
+  mutable bool keep_io_types_{false};
+  // float16 or bfloat16 now
+  mutable phi::DataType low_precision_{phi::DataType::FLOAT16};
+
+  mutable phi::Backend backend_{phi::Backend::GPU};
+
+  mutable std::unordered_set<std::string> black_list_;
+
+  // subgraph id -> pointer to subgraph
+  mutable std::vector<Graph*> subgraphes_;
+  // var name -> real var node
+  mutable std::unordered_map<std::string, Node*> real_vars_;
+  // subgraph id -> all op nodes in subgraph
+  mutable std::vector<std::vector<Node*>> all_op_nodes_;
+  // op's unique type -> the op's origin type
+  mutable std::unordered_map<std::string, std::string> op_original_type_;
+  // op's unique type -> whether the op run at low precision
+  mutable std::unordered_set<std::string> op_run_low_precision_;
+
+  mutable std::unordered_set<std::string> vars_convert_to_low_precision_;
+};
+
+bool OpSupportPrecision(const std::string& op_type,
+                        phi::Backend backend,
+                        phi::DataType precision,
+                        const std::unordered_set<std::string>& black_list);
+
+void DoInsertCastOp(Graph* graph,
+                    Node* var_node,
+                    Node* op_node,
+                    proto::VarType::Type from_type,
+                    proto::VarType::Type to_type,
+                    framework::BlockDesc* block_desc,
+                    int* suffix,
+                    std::unordered_map<Node*, Node*>* cache);
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/ir/delete_fill_constant_op_pass.cc

@@ -29,6 +29,11 @@ void FillConstData(LoDTensor* out_t, T value) {
 }
 
 void DeleteFillConstantOpPass::ApplyImpl(ir::Graph* graph) const {
+  bool with_dynamic_shape = Get<bool>("with_dynamic_shape");
+  // Not support
+  if (with_dynamic_shape) {
+    return;
+  }
   FusePassBase::Init("delete_fill_constant_op_pass", graph);
   GraphPatternDetector detector;
   auto fill_constant_op =

paddle/fluid/framework/ir/graph.cc

@@ -75,7 +75,6 @@ Graph::Graph(const ProgramDesc &program,
     }
   } else {
     auto var_nodes = InitFromProgram(program_, start_op_index, end_op_index);
-    ResolveHazard(var_nodes);
   }
 }
 
@@ -88,7 +87,6 @@ Graph::Graph(const BlockDesc &block,
              const int64_t end_op_index)
     : main_graph_(main_graph) {
   auto var_nodes = InitFromBlock(block, start_op_index, end_op_index);
-  ResolveHazard(var_nodes);
 }
 
 // TODO(levi): delete this interface after when we can convert all

paddle/fluid/framework/ir/graph_test.cc

@@ -130,86 +130,6 @@ TEST(GraphTest, Basic) {
   ASSERT_EQ(nodes.size(), 5UL);
 }
 
-TEST(GraphTest, WriteAfterRead) {
-  // void Test() {
-  ProgramDesc prog;
-  auto *op = prog.MutableBlock(0)->AppendOp();
-  op->SetType("sum");
-  op->SetInput("X", {"a"});
-  op->SetOutput("Out", {"b"});
-  op->SetAttr("op_role", 1);
-
-  op = prog.MutableBlock(0)->AppendOp();
-  op->SetType("dummy");
-  op->SetInput("X", {"c"});
-  op->SetOutput("Out", {"a"});
-  op->SetAttr("op_role", 1);
-
-  prog.MutableBlock(0)->Var("a")->SetType(proto::VarType::LOD_TENSOR);
-  prog.MutableBlock(0)->Var("b")->SetType(proto::VarType::LOD_TENSOR);
-  prog.MutableBlock(0)->Var("c")->SetType(proto::VarType::LOD_TENSOR);
-
-  std::unique_ptr<ir::Graph> g(new ir::Graph(prog));
-  ir::Node *control_dep1 = nullptr;
-  ir::Node *control_dep2 = nullptr;
-  for (ir::Node *n : g->Nodes()) {
-    if (n->Name() == "sum") {
-      ASSERT_EQ(n->outputs[0]->Name(), "b");
-      ASSERT_TRUE(ir::IsControlDepVar(*n->outputs[1]));
-      control_dep1 = n->outputs[1];
-      ASSERT_EQ(n->outputs.size(), 2UL);
-    }
-    if (n->Name() == "dummy") {
-      ASSERT_EQ(n->inputs[0]->Name(), "c");
-      ASSERT_TRUE(ir::IsControlDepVar(*n->inputs[1]));
-      control_dep2 = n->inputs[1];
-      ASSERT_EQ(n->inputs.size(), 2UL);
-    }
-  }
-  ASSERT_EQ(control_dep1, control_dep2);
-}
-
-TEST(GraphTest, WriteAfterWrite) {
-  // void Test() {
-  ProgramDesc prog;
-  auto *op = prog.MutableBlock(0)->AppendOp();
-  op->SetType("sum");
-  op->SetInput("X", {"a"});
-  op->SetOutput("Out", {"b"});
-  op->SetAttr("op_role", 1);
-
-  op = prog.MutableBlock(0)->AppendOp();
-  op->SetType("dummy");
-  op->SetInput("X", {"c"});
-  op->SetOutput("Out", {"b"});
-  op->SetAttr("op_role", 1);
-
-  prog.MutableBlock(0)->Var("a")->SetType(proto::VarType::LOD_TENSOR);
-  prog.MutableBlock(0)->Var("b")->SetType(proto::VarType::LOD_TENSOR);
-  prog.MutableBlock(0)->Var("c")->SetType(proto::VarType::LOD_TENSOR);
-
-  std::unique_ptr<ir::Graph> g(new ir::Graph(prog));
-  ir::Node *control_dep1 = nullptr;
-  ir::Node *control_dep2 = nullptr;
-  for (ir::Node *n : g->Nodes()) {
-    if (n->Name() == "sum") {
-      ASSERT_EQ(n->outputs[0]->Name(), "b");
-      ASSERT_TRUE(ir::IsControlDepVar(*n->outputs[1]));
-      ASSERT_EQ(n->outputs.size(), 2UL);
-      control_dep1 = n->outputs[1];
-    }
-    if (n->Name() == "dummy") {
-      ASSERT_EQ(n->inputs[0]->Name(), "c");
-      ASSERT_TRUE(ir::IsControlDepVar(*n->inputs[1]));
-      control_dep2 = n->inputs[1];
-      ASSERT_EQ(n->inputs.size(), 2UL);
-    }
-  }
-  ASSERT_NE(control_dep1, nullptr);
-  ASSERT_NE(control_dep2, nullptr);
-  ASSERT_EQ(control_dep1, control_dep2);
-}
-
 TEST(GraphTest, TestException) {
   ProgramDesc prog;
   std::unique_ptr<ir::Graph> g(new ir::Graph(prog));
@@ -350,12 +270,13 @@ TEST(GraphTest, TestMultiBlock) {
   op = prog.MutableBlock(1)->AppendOp();
   op->SetType("dummy");
   op->SetInput("X", {"c"});
-  op->SetOutput("Out", {"a"});
+  op->SetOutput("Out", {"d"});
   op->SetAttr("op_role", 1);
 
   prog.MutableBlock(1)->Var("a")->SetType(proto::VarType::LOD_TENSOR);
   prog.MutableBlock(1)->Var("b")->SetType(proto::VarType::LOD_TENSOR);
   prog.MutableBlock(1)->Var("c")->SetType(proto::VarType::LOD_TENSOR);
+  prog.MutableBlock(1)->Var("d")->SetType(proto::VarType::LOD_TENSOR);
 
   // Set contents in block_2.
   op = prog.MutableBlock(2)->AppendOp();
@@ -367,12 +288,13 @@ TEST(GraphTest, TestMultiBlock) {
   op = prog.MutableBlock(2)->AppendOp();
   op->SetType("dummy");
   op->SetInput("X", {"c"});
-  op->SetOutput("Out", {"b"});
+  op->SetOutput("Out", {"d"});
   op->SetAttr("op_role", 1);
 
   prog.MutableBlock(2)->Var("a")->SetType(proto::VarType::LOD_TENSOR);
   prog.MutableBlock(2)->Var("b")->SetType(proto::VarType::LOD_TENSOR);
   prog.MutableBlock(2)->Var("c")->SetType(proto::VarType::LOD_TENSOR);
+  prog.MutableBlock(1)->Var("d")->SetType(proto::VarType::LOD_TENSOR);
 
   // Step2: Convert program into graph, 3 blocks corresponding 3 sub_graphs.
   std::unique_ptr<ir::Graph> g(new ir::Graph(prog));
@@ -399,45 +321,29 @@ TEST(GraphTest, TestMultiBlock) {
 
   // Check contents in sub_graph_1.
   const ir::Graph *g1 = g->GetSubGraph(1);
-  ir::Node *control_dep1 = nullptr;
-  ir::Node *control_dep2 = nullptr;
   for (ir::Node *n : g1->Nodes()) {
     if (n->Name() == "sum") {
       ASSERT_EQ(n->outputs[0]->Name(), "b");
-      ASSERT_TRUE(ir::IsControlDepVar(*n->outputs[1]));
-      control_dep1 = n->outputs[1];
-      ASSERT_EQ(n->outputs.size(), 2UL);
+      ASSERT_EQ(n->outputs.size(), 1UL);
     }
     if (n->Name() == "dummy") {
       ASSERT_EQ(n->inputs[0]->Name(), "c");
-      ASSERT_TRUE(ir::IsControlDepVar(*n->inputs[1]));
-      control_dep2 = n->inputs[1];
-      ASSERT_EQ(n->inputs.size(), 2UL);
+      ASSERT_EQ(n->inputs.size(), 1UL);
     }
   }
-  ASSERT_EQ(control_dep1, control_dep2);
 
   // Check contents in sub_graph_2.
   const ir::Graph *g2 = g->GetSubGraph(2);
-  control_dep1 = nullptr;
-  control_dep2 = nullptr;
   for (ir::Node *n : g2->Nodes()) {
     if (n->Name() == "sum") {
       ASSERT_EQ(n->outputs[0]->Name(), "b");
-      ASSERT_TRUE(ir::IsControlDepVar(*n->outputs[1]));
-      ASSERT_EQ(n->outputs.size(), 2UL);
-      control_dep1 = n->outputs[1];
+      ASSERT_EQ(n->outputs.size(), 1UL);
     }
     if (n->Name() == "dummy") {
       ASSERT_EQ(n->inputs[0]->Name(), "c");
-      ASSERT_TRUE(ir::IsControlDepVar(*n->inputs[1]));
-      control_dep2 = n->inputs[1];
-      ASSERT_EQ(n->inputs.size(), 2UL);
+      ASSERT_EQ(n->inputs.size(), 1UL);
     }
   }
-  ASSERT_NE(control_dep1, nullptr);
-  ASSERT_NE(control_dep2, nullptr);
-  ASSERT_EQ(control_dep1, control_dep2);
 
   // Step3: Clone graph.
   std::shared_ptr<ir::Graph> clone_g = g->Clone();

paddle/fluid/framework/ir/multi_batch_merge_pass.cc

@@ -331,8 +331,6 @@ void BatchMergePass::ApplyImpl(ir::Graph* graph) const {
       copy_node(node);
     }
   }
-
-  result.ResolveHazard(created);
 }
 
 }  // namespace ir

paddle/fluid/framework/naive_executor.cc

@@ -183,5 +183,6 @@ void NaiveExecutor::ResetTrtOps(int num) {
   }
 #endif
 }
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/inference/analysis/analyzer.cc

@@ -38,8 +38,7 @@ void Analyzer::RunAnalysis(Argument *argument) {
     if (!disable_logs) {
       string::PrettyLogH1("--- Running analysis [%s]", pass);
     }
-    if (!argument->enable_analysis_optim() && pass == "ir_analysis_pass")
-      continue;
+    if (!argument->enable_ir_optim() && pass == "ir_analysis_pass") continue;
 
     auto *ptr = PassRegistry::Global().Retreive(pass);
     PADDLE_ENFORCE_NOT_NULL(ptr,

paddle/fluid/inference/analysis/analyzer_tester.cc

@@ -31,7 +31,7 @@ TEST(Analyzer, analysis_without_tensorrt) {
   Argument argument;
   argument.SetDisableLogs(false);
   argument.SetModelDir(FLAGS_inference_model_dir);
-  argument.SetEnableAnalysisOptim(false);
+  argument.SetEnableIrOptim(false);
   argument.SetUseGPU(false);
   argument.SetAnalysisPasses({"ir_graph_build_pass",
                               "ir_analysis_pass",
@@ -44,7 +44,7 @@ TEST(Analyzer, analysis_without_tensorrt) {
 TEST(Analyzer, analysis_with_tensorrt) {
   Argument argument;
   argument.SetDisableLogs(false);
-  argument.SetEnableAnalysisOptim(false);
+  argument.SetEnableIrOptim(false);
   argument.SetTensorRtMaxBatchSize(3);
   argument.SetTensorRtWorkspaceSize(1 << 20);
   argument.SetModelDir(FLAGS_inference_model_dir);

paddle/fluid/inference/analysis/argument.h

@@ -42,8 +42,6 @@ namespace paddle {
 namespace inference {
 namespace analysis {
 
-using framework::ir::Graph;
-
 #ifdef PADDLE_WITH_MKLDNN
 using VarQuantScale =
     std::unordered_map<std::string, std::pair<bool, framework::LoDTensor>>;
@@ -148,7 +146,7 @@ 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);
+  DECL_ARGUMENT_FIELD(enable_ir_optim, EnableIrOptim, bool);
 
   // For JITLayer
   DECL_ARGUMENT_FIELD(skip_load_params, SkipLoadParams, bool);
@@ -362,6 +360,8 @@ struct Argument {
   DECL_ARGUMENT_FIELD(mixed_black_list,
                       MixedBlackList,
                       std::unordered_set<std::string>);
+  DECL_ARGUMENT_FIELD(enable_gpu_mixed, EnableGPUMixed, bool);
+  DECL_ARGUMENT_FIELD(mixed_precision_mode, MixedPrecisionMode, int);
 
  private:
   std::unordered_set<std::string> valid_fields_;

paddle/fluid/inference/analysis/helper.h

@@ -153,25 +153,6 @@ T &GetFromScope(const framework::Scope &scope, const std::string &name) {
   return *var->GetMutable<T>();
 }
 
-static framework::proto::ProgramDesc LoadProgramDesc(
-    const std::string &model_path) {
-  std::ifstream fin(model_path, std::ios::in | std::ios::binary);
-  PADDLE_ENFORCE_EQ(
-      fin.is_open(),
-      true,
-      platform::errors::NotFound(
-          "Cannot open file %s, please confirm whether the file exists",
-          model_path));
-  fin.seekg(0, std::ios::end);
-  std::string buffer(fin.tellg(), ' ');
-  fin.seekg(0, std::ios::beg);
-  fin.read(&buffer[0], buffer.size());
-  fin.close();
-  framework::proto::ProgramDesc program_desc;
-  program_desc.ParseFromString(buffer);
-  return program_desc;
-}
-
 static bool FileExists(const std::string &filepath) {
   std::ifstream file(filepath);
   bool exists = file.is_open();

paddle/fluid/inference/analysis/ir_pass_manager.cc

@@ -27,6 +27,7 @@
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/inference/analysis/argument.h"
 #include "paddle/fluid/string/pretty_log.h"
+#include "paddle/phi/core/errors.h"
 
 namespace paddle {
 namespace inference {
@@ -36,15 +37,6 @@ using string::PrettyLogEndl;
 using string::Style;
 
 IRPassManager::IRPassManager(Argument *argument) {
-  ARGUMENT_CHECK_FIELD(argument, main_program);
-  graph_ = std::unique_ptr<Graph>(new Graph(argument->main_program()));
-  if (argument->Has("scope")) {
-    auto *scope_ptr = argument->scope_ptr();
-    PADDLE_ENFORCE_NOT_NULL(scope_ptr,
-                            platform::errors::PreconditionNotMet(
-                                "The scope ptr should not be nullptr."));
-    graph_->SetNotOwned(framework::ir::kParamScopeAttr, scope_ptr);
-  }
   disable_logs_ = argument->disable_logs();
 
   ARGUMENT_CHECK_FIELD(argument, ir_analysis_passes);
@@ -95,10 +87,14 @@ void IRPassManager::CreatePasses(Argument *argument,
                               argument->tensorrt_tuned_dynamic_shape();
     pass->Set("with_dynamic_shape", new bool(with_dynamic_shape));
 
+    // mixed precision related
     pass->Set("model_precision", new int(argument->model_precision()));
     pass->Set(
         "mixed_black_list",
         new std::unordered_set<std::string>(argument->mixed_black_list()));
+    pass->Set("enable_gpu_mixed", new bool(argument->enable_gpu_mixed()));
+    pass->Set("mixed_precision_mode",
+              new int(argument->mixed_precision_mode()));
 
     if (pass_name == "graph_viz_pass") {
       std::string optim_cache_dir = argument->optim_cache_dir();
@@ -302,42 +298,18 @@ void IRPassManager::CreatePasses(Argument *argument,
 }
 
 std::unique_ptr<Graph> IRPassManager::Apply(std::unique_ptr<Graph> graph) {
-  if (passes_.empty()) {
-    return graph;
-  }
   PADDLE_ENFORCE_NOT_NULL(
-      graph.get(),
-      platform::errors::PreconditionNotMet("Graph cannot be NULL."));
+      graph.get(), platform::errors::InvalidArgument("Graph cannot be null."));
   // Apply all the passes
   for (const auto &pass : passes_) {
     if (pass->Type() != "graph_viz_pass" && !disable_logs_) {
       PrettyLogEndl(Style::H2(), "--- Running IR pass [%s]", pass->Type());
     }
-    // delete_fill_constant_op_pass is not apply under trt dynamic shape
-    if (pass->Type() == "delete_fill_constant_op_pass") {
-      bool use_dynamic = pass->Get<bool>("with_dynamic_shape");
-      if (use_dynamic) continue;
-    }
     graph.reset(pass->Apply(graph.release()));
   }
   return graph;
 }
 
-framework::proto::ProgramDesc IRPassManager::AcquireProgram(
-    std::unique_ptr<Graph> *graph, ProgramDesc *program) const {
-  auto pass =
-      framework::ir::PassRegistry::Instance().Get("graph_to_program_pass");
-
-  // Direct using ProgramDesc desc(argument->main_program()) may cause
-  // incomplete copies of information.
-  ProgramDesc desc;
-  desc.CopyFrom(*program->Proto());
-  pass->SetNotOwned("program", &desc);
-  auto *the_graph = graph->release();
-  graph->reset(pass->Apply(the_graph));
-  return *desc.Proto();
-}
-
 }  // namespace analysis
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/analysis/ir_pass_manager.h

@@ -48,15 +48,9 @@ class IRPassManager final {
 
   std::unique_ptr<Graph> Apply(std::unique_ptr<Graph> graph);
 
-  framework::proto::ProgramDesc AcquireProgram(std::unique_ptr<Graph> *graph,
-                                               ProgramDesc *program) const;
-
-  framework::ir::Graph &graph() const { return *graph_; }
-
  private:
   void CreatePasses(Argument *argument, const std::vector<std::string> &passes);
 
-  std::unique_ptr<Graph> graph_;
   std::vector<std::unique_ptr<Pass>> passes_;
   bool disable_logs_{false};
 };

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

@@ -94,14 +94,14 @@ void OutputProcess(framework::ir::Graph *graph,
                   backend,
                   precision,
                   blacklist)) {
-            AddCastOp(graph,
-                      var_node,
-                      next_op,
-                      framework::proto::VarType::FP32,
-                      to_type,
-                      &suffix,
-                      block_desc,
-                      &var_to_cast_op_map);
+            InsertCastOp(graph,
+                         var_node,
+                         next_op,
+                         framework::proto::VarType::FP32,
+                         to_type,
+                         block_desc,
+                         &suffix,
+                         &var_to_cast_op_map);
             var_node->Var()->SetDataType(framework::proto::VarType::FP32);
           }
         }

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

@@ -13,7 +13,7 @@ cc_library(
 cc_library(
   convert_to_mixed_precision
   SRCS convert_to_mixed_precision.cc
-  DEPS analysis_pass ir_graph_build_pass)
+  DEPS analysis_pass ir_graph_build_pass auto_mixed_precision_pass)
 cc_library(
   ir_params_sync_among_devices_pass
   SRCS ir_params_sync_among_devices_pass.cc
@@ -30,17 +30,6 @@ cc_library(
   inference_op_replace_pass
   SRCS inference_op_replace_pass.cc
   DEPS analysis_pass graph_to_program_pass)
-if(WITH_TESTING)
-  cc_library(
-    ir_graph_clean_pass
-    SRCS ir_graph_clean_pass.cc
-    DEPS analysis_pass gtest)
-else()
-  cc_library(
-    ir_graph_clean_pass
-    SRCS ir_graph_clean_pass.cc
-    DEPS analysis_pass)
-endif()
 
 cc_library(
   analysis_passes
@@ -52,8 +41,7 @@ cc_library(
        memory_optim_pass
        convert_to_mixed_precision
        inference_op_replace_pass
-       ir_graph_to_program_pass
-       ir_graph_clean_pass)
+       ir_graph_to_program_pass)
 
 set(analysis_deps
     ${analysis_deps} analysis_passes subgraph_detector

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

@@ -15,14 +15,12 @@
 #pragma once
 
 #include <string>
-#include <unordered_map>
 #include <unordered_set>
 
 #include "paddle/fluid/framework/block_desc.h"
 #include "paddle/fluid/framework/ir/graph.h"
-#include "paddle/fluid/framework/ir/graph_helper.h"
-#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
 #include "paddle/fluid/framework/program_desc.h"
+#include "paddle/fluid/framework/scope.h"
 #include "paddle/phi/common/backend.h"
 #include "paddle/phi/common/data_type.h"
 
@@ -30,20 +28,52 @@ namespace paddle {
 namespace inference {
 namespace analysis {
 
+class ConvertToMixedPrecisionPass {
+ public:
+  explicit ConvertToMixedPrecisionPass(
+      const std::string& model_file,
+      const std::string& params_file,
+      const std::string& mixed_model_file,
+      const std::string& mixed_params_file,
+      phi::DataType mixed_precision,
+      phi::Backend backend,
+      bool keep_io_types,
+      const std::unordered_set<std::string>& black_list);
+
+  void Run();
+
+ private:
+  void LoadModel();
+  void SaveMixedModel();
+
+ private:
+  std::string model_file_;
+  std::string params_file_;
+  std::string mixed_model_file_;
+  std::string mixed_params_file_;
+  phi::DataType mixed_precision_;
+  phi::Backend backend_;
+  bool keep_io_types_;
+  std::unordered_set<std::string> black_list_;
+
+  framework::Scope scope_;
+  std::unique_ptr<framework::ir::Graph> main_graph_{nullptr};
+};
+
 bool OpSupportPrecision(const std::string& op_type,
                         phi::Backend backend,
                         phi::DataType precision,
-                        const std::unordered_set<std::string>& blacklist);
+                        const std::unordered_set<std::string>& black_list);
 
-void AddCastOp(
+void InsertCastOp(
     framework::ir::Graph* graph,
-    framework::ir::Node* node,
-    framework::ir::Node* next_op,
+    framework::ir::Node* var_node,
+    framework::ir::Node* op_node,
     framework::proto::VarType::Type from_type,
     framework::proto::VarType::Type to_type,
-    int* suffix,
     framework::BlockDesc* block_desc,
-    std::unordered_map<framework::ir::Node*, framework::ir::Node*>* map);
+    int* suffix,
+    std::unordered_map<framework::ir::Node*, framework::ir::Node*>* visited);
 
 void ConvertToMixedPrecision(const std::string& model_file,
                              const std::string& params_file,
@@ -51,8 +81,8 @@ void ConvertToMixedPrecision(const std::string& model_file,
                              const std::string& mixed_params_file,
                              phi::DataType mixed_precision,
                              phi::Backend backend,
-                             bool keep_io_types = true,
-                             std::unordered_set<std::string> black_list = {});
+                             bool keep_io_types,
+                             const std::unordered_set<std::string>& black_list);
 
 }  // namespace analysis
 }  // namespace inference

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

@@ -40,7 +40,7 @@ void InferenceOpReplacePass::RunImpl(Argument* argument) {
 }
 
 std::string InferenceOpReplacePass::repr() const {
-  return "inference-op-replace-pass";
+  return "inference_op_replace_pass";
 }
 
 }  // namespace analysis

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

@@ -105,7 +105,7 @@ void IrAnalysisPass::CollectFusionStatis(Argument* argument) {
           framework::ir::kFuseStatisAttr));
 }
 
-std::string IrAnalysisPass::repr() const { return "ir-analysis-pass"; }
+std::string IrAnalysisPass::repr() const { return "ir_analysis_pass"; }
 
 }  // namespace analysis
 }  // namespace inference

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

@@ -64,7 +64,8 @@ void IrGraphBuildPass::RunImpl(Argument *argument) {
         "set."));
   }
 
-  auto graph = std::unique_ptr<Graph>(new Graph(argument->main_program()));
+  auto graph = std::unique_ptr<framework::ir::Graph>(
+      new framework::ir::Graph(argument->main_program()));
   argument->SetMainGraph(graph.release());
   auto *scope_ptr = argument->scope_ptr();
   PADDLE_ENFORCE_NOT_NULL(scope_ptr,
@@ -125,7 +126,7 @@ std::unique_ptr<framework::ProgramDesc> IrGraphBuildPass::LoadModel(
   }
 }
 
-std::string IrGraphBuildPass::repr() const { return "ir-graph-build-pass"; }
+std::string IrGraphBuildPass::repr() const { return "ir_graph_build_pass"; }
 
 }  // namespace analysis
 }  // namespace inference

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

-// 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.
-
-#include "paddle/fluid/inference/analysis/passes/ir_graph_clean_pass.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 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,
-                            platform::errors::PreconditionNotMet(
-                                "The node should not be nullptr."));
-    if (is_valid_node(node)) {
-      invalid_nodes.insert(node);
-    } else if (node->IsOp()) {
-      ++valid_op;
-    }
-  }
-
-  GraphSafeRemoveNodes(&graph, invalid_nodes);
-}
-
-}  // namespace analysis
-}  // namespace inference
-}  // namespace paddle

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 {
-
-struct Argument;
-
-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/ir_graph_to_program_pass.cc

@@ -31,7 +31,7 @@ void IrGraphToProgramPass::RunImpl(Argument *argument) {
               new int(argument->memory_optim_sort_kind()));
   }
 
-  std::unique_ptr<Graph> graph(argument->main_graph_ptr());
+  std::unique_ptr<framework::ir::Graph> graph(argument->main_graph_ptr());
 
   // Direct using ProgramDesc desc(argument->main_program()) may cause
   // incomplete copies of information.

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

@@ -28,7 +28,7 @@ class IrGraphToProgramPass : public AnalysisPass {
  public:
   void RunImpl(Argument *argument) override;
 
-  std::string repr() const override { return "ir-graph-to-param-pass"; }
+  std::string repr() const override { return "ir_graph_to_param_pass"; }
 };
 
 }  // namespace analysis

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

@@ -169,7 +169,7 @@ void IrParamsSyncAmongDevicesPass::RunImpl(Argument *argument) {
 }
 
 std::string IrParamsSyncAmongDevicesPass::repr() const {
-  return "ir-params-sync-among-devices-pass";
+  return "ir_params_sync_among_devices_pass";
 }
 
 }  // namespace analysis

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

@@ -295,7 +295,7 @@ void UpdateOpDescsByReuse(
   }
 }
 
-std::string MemoryOptimizePass::repr() const { return "memory optimize pass"; }
+std::string MemoryOptimizePass::repr() const { return "memory_optimize_pass"; }
 
 void MemoryOptimizePass::RunImpl(Argument* argument) {
   // Memory optimization.

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

@@ -18,7 +18,6 @@
 #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"
@@ -34,8 +33,6 @@ 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_config.cc

@@ -85,15 +85,29 @@ void AnalysisConfig::SetModel(const std::string &prog_file_path,
 
   Update();
 }
+
 void AnalysisConfig::EnableUseGpu(uint64_t memory_pool_init_size_mb,
-                                  int device_id) {
+                                  int device_id,
+                                  Precision precision_mode) {
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
   use_gpu_ = true;
   memory_pool_init_size_mb_ = memory_pool_init_size_mb;
   FLAGS_initial_gpu_memory_in_mb = memory_pool_init_size_mb_;
   gpu_device_id_ = device_id;
+  mixed_precision_mode_ = precision_mode;
+  if (precision_mode == Precision::kFloat32) {
+    // default
+  } else if (precision_mode == Precision::kHalf ||
+             precision_mode == Precision::kBf16) {
+    enable_gpu_mixed_ = true;
+  } else {
+    LOG(ERROR)
+        << "The Paddle-GPU inference currently only supports "
+           "float32/float16/bfloat16 precision. Please check the parameters "
+           "you specified in EnableUseGpu or enable_use_gpu function.";
+  }
 #else
-  LOG(ERROR) << "Please compile with gpu to EnableGpu()";
+  LOG(ERROR) << "Please use PaddlePaddle with GPU version.";
   use_gpu_ = false;
 #endif
 
@@ -279,7 +293,7 @@ void AnalysisConfig::LoadIpuConfig(const std::string &config_path) {
 
     if (ipu_config_mapper_.find(key) == ipu_config_mapper_.end()) {
       PADDLE_THROW(platform::errors::InvalidArgument(
-          "invalid key {} in IPU config", key));
+          "invalid key %s in IPU config: ", key));
     }
     switch (ipu_config_mapper_.at(key)) {
       case ipu_config_code::ipu_device_num:
@@ -315,7 +329,7 @@ void AnalysisConfig::LoadIpuConfig(const std::string &config_path) {
 
       default:
         PADDLE_THROW(platform::errors::InvalidArgument(
-            "invalid key {} in IPU config", key));
+            "invalid key %s in IPU config", key));
         break;
     }
   }
@@ -372,8 +386,10 @@ AnalysisConfig::AnalysisConfig(const AnalysisConfig &other) {
   CP_MEMBER(gpu_device_id_);
   CP_MEMBER(memory_pool_init_size_mb_);
 
-  // Mixed related.
+  // Mixed precision related.
   CP_MEMBER(mixed_black_list_);
+  CP_MEMBER(enable_gpu_mixed_);
+  CP_MEMBER(mixed_precision_mode_);
 
   CP_MEMBER(enable_memory_optim_);
   // TensorRT related.
@@ -740,13 +756,7 @@ void AnalysisConfig::Update() {
       ((use_custom_device() ^ pass_builder_->use_custom_device()))) {
     if (use_gpu()) {
       pass_builder_.reset(new GpuPassStrategy);
-
-      if (use_tensorrt_) {
-        // Append after the Affine_channel_conv_fuse pass.
-        pass_builder()->InsertPass(3, "tensorrt_subgraph_pass");
-      }
     } else if (use_ipu()) {
-      VLOG(1) << "IpuPassStrategy has been used for new.";
       pass_builder_.reset(new IpuPassStrategy);
     } else if (use_xpu()) {
       PADDLE_ENFORCE_EQ(
@@ -946,9 +956,6 @@ void AnalysisConfig::Update() {
         "but did not have the option -DWITH_CUSTOM_DEVICE compiled."));
 #endif
   }
-  if (ir_debug_) {
-    pass_builder()->TurnOnDebug();
-  }
 }
 
 std::string AnalysisConfig::SerializeInfoCache() {
@@ -960,6 +967,7 @@ std::string AnalysisConfig::SerializeInfoCache() {
   ss << calibration_file_path_;
 
   ss << use_gpu_;
+  ss << enable_gpu_mixed_;
   ss << use_external_stream_;
   ss << exec_stream_;
   ss << use_fc_padding_;
@@ -1167,6 +1175,7 @@ std::string AnalysisConfig::Summary() {
   os.InsertRow({"use_gpu", use_gpu_ ? "true" : "false"});
   if (use_gpu_) {
     os.InsertRow({"gpu_device_id", std::to_string(gpu_device_id_)});
+    os.InsertRow({"enable_gpu_mixed", std::to_string(enable_gpu_mixed_)});
     os.InsertRow({"memory_pool_init_size",
                   std::to_string(memory_pool_init_size_mb_) + "MB"});
     os.InsertRow(
@@ -1360,7 +1369,7 @@ bool AnalysisConfig::trt_allow_build_at_runtime() {
   return trt_allow_build_at_runtime_;
 }
 
-void AnalysisConfig::Exp_SetBlackListOpsForMixedModel(
+void AnalysisConfig::Exp_DisableMixedPrecisionOps(
     const std::unordered_set<std::string> &black_list) {
   mixed_black_list_ = black_list;
 }

paddle/fluid/inference/api/analysis_predictor.cc

@@ -1065,7 +1065,7 @@ void AnalysisPredictor::PrepareArgument() {
   argument_.SetUseGPU(config_.use_gpu());
   argument_.SetUseFcPadding(config_.use_fc_padding());
   argument_.SetGPUDeviceId(config_.gpu_device_id());
-  argument_.SetEnableAnalysisOptim(config_.enable_ir_optim_);
+  argument_.SetEnableIrOptim(config_.enable_ir_optim_);
   argument_.SetEnableMemoryOptim(config_.enable_memory_optim());
   argument_.SetModelFromMemory(config_.model_from_memory_);
   // Analyze inference_program
@@ -1210,53 +1210,57 @@ void AnalysisPredictor::PrepareArgument() {
   }
 #endif
 
-  auto passes = config_.pass_builder()->AllPasses();
+  auto *pass_builder = config_.pass_builder();
   if (model_precision_ != phi::DataType::FLOAT32) {
     LOG(INFO) << "Model is mixed precision type with " << model_precision_
               << ", we will use a new PassStrategy. Note that only the GPU "
                  "backend is supported for now.";
-    passes.clear();
+    pass_builder->ClearPasses();
+    const auto &deleted_passes = pass_builder->GetAllDeletedPasses();
     if (config_.tensorrt_engine_enabled()) {
       for (const auto &pass : kTrtLowerPrecisionPasses) {
-        passes.push_back(pass);
+        if (deleted_passes.count(pass)) continue;
+        pass_builder->AppendPass(pass);
       }
     } else if (config_.use_gpu()) {
       for (const auto &pass : kGpuLowerPrecisionPasses) {
-        passes.push_back(pass);
+        if (deleted_passes.count(pass)) continue;
+        pass_builder->AppendPass(pass);
       }
     }
+  }
 
-    const auto &deleted_passes = config_.pass_builder()->GetAllDeletedPasses();
-    for (const auto &it : deleted_passes) {
-      auto iterator = std::find(passes.begin(), passes.end(), it);
-      if (iterator != passes.end()) {
-        passes.erase(iterator);
-      }
+  if (!config_.ir_optim()) {
+    argument_.SetEnableIrOptim(false);
+    if (config_.enable_gpu_mixed_) {
+      argument_.SetEnableIrOptim(true);
+      pass_builder->ClearPasses();
+      pass_builder->AppendPass("auto_mixed_precision_pass");
+      LOG(INFO)
+          << "This model run in Paddle-GPU mixed precision mode with no ir "
+             "optimization.";
+    } else {
+      LOG(INFO) << "ir_optim is turned off, no IR pass will be executed.";
     }
-
+  } else {
     if (config_.ir_debug_) {
-      auto it = std::begin(passes);
-      while (it != std::end(passes)) {
-        if (*it != "graph_viz_pass") {
-          it = passes.insert(it + 1, "graph_viz_pass");
-        } else {
-          ++it;
-        }
-      }
+      pass_builder->TurnOnDebug();
+    }
+    if (config_.enable_gpu_mixed_) {
+      LOG(INFO) << "This model run in Paddle-GPU mixed precision mode.";
     }
-  }
-  if (!config_.ir_optim()) {
-    passes.clear();
-    LOG(INFO) << "ir_optim is turned off, no IR pass will be executed";
   }
   argument_.SetDisableLogs(config_.glog_info_disabled());
-  argument_.SetIrAnalysisPasses(passes);
-  argument_.SetAnalysisPasses(config_.pass_builder()->AnalysisPasses());
+  argument_.SetIrAnalysisPasses(pass_builder->AllPasses());
+  argument_.SetAnalysisPasses(pass_builder->AnalysisPasses());
   argument_.SetScopeNotOwned(scope_.get());
 
   // mixed precison.
   argument_.SetModelPrecision(static_cast<int>(model_precision_));
   argument_.SetMixedBlackList(config_.mixed_black_list_);
+  argument_.SetEnableGPUMixed(config_.enable_gpu_mixed_);
+  argument_.SetMixedPrecisionMode(static_cast<int>(
+      paddle::ConvertPrecision(config_.mixed_precision_mode_)));
 }
 
 // NOTE All the members in AnalysisConfig should be copied to Argument.
@@ -2107,7 +2111,9 @@ std::unique_ptr<PaddlePredictor> AnalysisPredictor::Clone(void *stream) {
   }
   x->predictor_stream_ = stream;
   x->Init(scope_, inference_program_);
+#ifdef PADDLE_WITH_TENSORRT
   x->executor_->ResetTrtOps(++AnalysisPredictor::clone_num_);
+#endif
   return std::unique_ptr<PaddlePredictor>(x);
 }
 

paddle/fluid/inference/api/mkldnn_quantizer.cc

@@ -604,10 +604,8 @@ void AnalysisPredictor::MkldnnQuantizer::PrepareArgument() const {
   if (predictor_.config_.ir_debug_) builder->TurnOnDebug();
   auto passes = builder->AllPasses();
   predictor_.argument_.SetIrAnalysisPasses(passes);
-  predictor_.argument_.SetAnalysisPasses({"ir_graph_clean_pass",
-                                          "ir_analysis_pass",
-                                          "memory_optimize_pass",
-                                          "ir_graph_to_program_pass"});
+  predictor_.argument_.SetAnalysisPasses(
+      {"ir_analysis_pass", "memory_optimize_pass", "ir_graph_to_program_pass"});
   predictor_.argument_.SetQuantVarScales(scales_);
 }
 

paddle/fluid/inference/api/paddle_analysis_config.h

@@ -247,8 +247,12 @@ struct PD_INFER_DECL AnalysisConfig {
   ///
   /// \param memory_pool_init_size_mb initial size of the GPU memory pool in MB.
   /// \param device_id device_id the GPU card to use (default is 0).
+  /// \param precision the precision used in Paddle-GPU inference.
   ///
-  void EnableUseGpu(uint64_t memory_pool_init_size_mb, int device_id = 0);
+  void EnableUseGpu(uint64_t memory_pool_init_size_mb,
+                    int device_id = 0,
+                    Precision precision_mode = Precision::kFloat32);
+
   ///
   /// \brief Turn off GPU.
   ///
@@ -967,7 +971,7 @@ struct PD_INFER_DECL AnalysisConfig {
   /// interface is in the experimental stage and may change in the future. Note
   /// that the blacklist must be the same as the model conversion blacklist.
   ///
-  void Exp_SetBlackListOpsForMixedModel(
+  void Exp_DisableMixedPrecisionOps(
       const std::unordered_set<std::string>& black_list);
 
   void SetApplyOptim(bool value) { apply_optim_ = value; }
@@ -987,13 +991,15 @@ struct PD_INFER_DECL AnalysisConfig {
   mutable std::string params_file_;
   mutable std::string calibration_file_path_;
 
-  // Mixed precision.
+  // Mixed precision related.
+  Precision mixed_precision_mode_{Precision::kFloat32};
   std::unordered_set<std::string> mixed_black_list_;
 
   // GPU related.
   bool use_gpu_{false};
   int gpu_device_id_{0};
   uint64_t memory_pool_init_size_mb_{100};  // initial size is 100MB.
+  bool enable_gpu_mixed_{false};
   bool thread_local_stream_{false};
 
   bool use_cudnn_{false};

paddle/fluid/inference/api/paddle_pass_builder.cc

@@ -227,9 +227,10 @@ GpuPassStrategy::GpuPassStrategy() : PassStrategy({}) {
         "conv_elementwise_add_fuse_pass",      //
 #endif                                         //
         "transpose_flatten_concat_fuse_pass",  //
-        "constant_folding_pass",
+        "constant_folding_pass",               //
         // following pass should be located in the last, since it will
         // work on all fused ops.
+        "auto_mixed_precision_pass",  //
         "runtime_context_cache_pass"
   });
 

paddle/fluid/inference/api/paddle_pass_builder.h

@@ -115,7 +115,6 @@ class PD_INFER_DECL PaddlePassBuilder {
   /// \cond Protected
   std::vector<std::string> analysis_passes_{
       {"ir_graph_build_pass",
-       "ir_graph_clean_pass",
        "ir_analysis_pass",
        "ir_params_sync_among_devices_pass",
        "adjust_cudnn_workspace_size_pass",

paddle/fluid/inference/tensorrt/engine.h

@@ -294,15 +294,6 @@ class TensorRTEngine {
 
   nvinfer1::ICudaEngine* engine() { return infer_engine_.get(); }
   nvinfer1::IExecutionContext* context() {
-#ifndef PADDLE_WITH_TESTING
-    PADDLE_ENFORCE_GT(
-        predictor_id_per_thread,
-        -1,
-        platform::errors::InvalidArgument(
-            "thread local var predictor_id_per_thread must be "
-            "initialized to >= 0, but now predictor_id_per_thread = %d",
-            predictor_id_per_thread));
-#endif
     std::unique_lock<std::mutex> lock(mutex_);
     if (infer_context_.find(predictor_id_per_thread) == infer_context_.end()) {
       PADDLE_ENFORCE_NOT_NULL(
@@ -329,15 +320,6 @@ class TensorRTEngine {
 
   int GetProfileIndex() {
     if (max_profile_num_ > 1) {
-#ifndef PADDLE_WITH_TESTING
-      PADDLE_ENFORCE_GT(
-          predictor_id_per_thread,
-          -1,
-          platform::errors::InvalidArgument(
-              "thread local var predictor_id_per_thread must be "
-              "initialized to >= 0, but now predictor_id_per_thread = %d",
-              predictor_id_per_thread));
-#endif
       std::unique_lock<std::mutex> lock(mutex_);
       return profile_index_[predictor_id_per_thread];
     } else {
@@ -356,15 +338,6 @@ class TensorRTEngine {
         infer_engine_,
         platform::errors::InvalidArgument(
             "You should build engine first and then set the context."));
-#ifndef PADDLE_WITH_TESTING
-    PADDLE_ENFORCE_GT(
-        predictor_id_per_thread,
-        -1,
-        platform::errors::InvalidArgument(
-            "thread local var predictor_id_per_thread must be "
-            "initialized to >= 0, but now predictor_id_per_thread = %d",
-            predictor_id_per_thread));
-#endif
     std::unique_lock<std::mutex> lock(mutex_);
     infer_context_[predictor_id_per_thread].reset(nullptr);
     infer_context_.erase(predictor_id_per_thread);

paddle/fluid/inference/tests/api/CMakeLists.txt

@@ -416,6 +416,9 @@ download_result(${ERNIE_INSTALL_DIR} "Ernie_result.txt.tar.gz"
 if(WITH_GPU)
   inference_analysis_api_test(test_analyzer_ernie ${ERNIE_INSTALL_DIR}
                               analyzer_ernie_tester.cc)
+  inference_analysis_api_test(gpu_ernie_half_test ${ERNIE_INSTALL_DIR}
+                              gpu_ernie_half_test.cc)
+  set_tests_properties(gpu_ernie_half_test PROPERTIES TIMEOUT 60)
 endif()
 inference_analysis_api_int8_test(test_analyzer_ernie_int8 ${ERNIE_INSTALL_DIR}
                                  analyzer_ernie_int8_tester.cc)

paddle/fluid/inference/tests/api/gpu_ernie_half_test.cc

+// Copyright (c) 2022 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.
+
+#include "paddle/fluid/inference/api/paddle_inference_api.h"
+#include "paddle/fluid/inference/tests/api/tester_helper.h"
+
+namespace paddle {
+namespace inference {
+
+using paddle::PaddleTensor;
+
+template <typename T>
+void GetValueFromStream(std::stringstream *ss, T *t) {
+  (*ss) >> (*t);
+}
+
+template <>
+void GetValueFromStream<std::string>(std::stringstream *ss, std::string *t) {
+  *t = ss->str();
+}
+
+// Split string to vector
+template <typename T>
+void Split(const std::string &line, char sep, std::vector<T> *v) {
+  std::stringstream ss;
+  T t;
+  for (auto c : line) {
+    if (c != sep) {
+      ss << c;
+    } else {
+      GetValueFromStream<T>(&ss, &t);
+      v->push_back(std::move(t));
+      ss.str({});
+      ss.clear();
+    }
+  }
+
+  if (!ss.str().empty()) {
+    GetValueFromStream<T>(&ss, &t);
+    v->push_back(std::move(t));
+    ss.str({});
+    ss.clear();
+  }
+}
+
+// Parse tensor from string
+template <typename T>
+bool ParseTensor(const std::string &field, paddle::PaddleTensor *tensor) {
+  std::vector<std::string> data;
+  Split(field, ':', &data);
+  if (data.size() < 2) return false;
+
+  std::string shape_str = data[0];
+
+  std::vector<int> shape;
+  Split(shape_str, ' ', &shape);
+
+  std::string mat_str = data[1];
+
+  std::vector<T> mat;
+  Split(mat_str, ' ', &mat);
+
+  tensor->shape = shape;
+  auto size =
+      std::accumulate(shape.begin(), shape.end(), 1, std::multiplies<int>()) *
+      sizeof(T);
+  tensor->data.Resize(size);
+  std::copy(mat.begin(), mat.end(), static_cast<T *>(tensor->data.data()));
+  tensor->dtype = GetPaddleDType<T>();
+
+  return true;
+}
+
+// Parse input tensors from string
+bool ParseLine(const std::string &line,
+               std::vector<paddle::PaddleTensor> *tensors) {
+  std::vector<std::string> fields;
+  Split(line, ';', &fields);
+
+  tensors->clear();
+  tensors->reserve(4);
+
+  int i = 0;
+  auto input_name = FLAGS_ernie_large ? "eval_placeholder_" : "placeholder_";
+  for (; i < 3; i++) {
+    paddle::PaddleTensor temp;
+    ParseTensor<int64_t>(fields[i], &temp);
+    temp.name = input_name + std::to_string(i);
+    tensors->push_back(temp);
+  }
+
+  // input_mask
+  paddle::PaddleTensor input_mask;
+  ParseTensor<float>(fields[i], &input_mask);
+  input_mask.name = input_name + std::to_string(i);
+  tensors->push_back(input_mask);
+
+  return true;
+}
+
+bool LoadInputData(std::vector<std::vector<paddle::PaddleTensor>> *inputs,
+                   int batch_size = 1) {
+  if (FLAGS_infer_data.empty()) {
+    LOG(ERROR) << "please set input data path";
+    return false;
+  }
+
+  std::ifstream fin(FLAGS_infer_data);
+  std::string line;
+  int sample = 0;
+
+  // The unit-test dataset only have 10 samples, each sample have 5 feeds.
+  while (std::getline(fin, line)) {
+    std::vector<paddle::PaddleTensor> feed_data;
+    ParseLine(line, &feed_data);
+    inputs->push_back(std::move(feed_data));
+    sample++;
+    if (!FLAGS_test_all_data && sample == batch_size) break;
+  }
+  LOG(INFO) << "number of samples: " << sample;
+  return true;
+}
+
+// Compare results
+TEST(Ernie_gpu_fp16_no_ir, compare_results) {
+  AnalysisConfig config;
+  config.SetModel(FLAGS_infer_model);
+  config.EnableUseGpu(512, 0, paddle_infer::PrecisionType::kHalf);
+  config.SwitchIrOptim(false);
+
+  auto predictor = CreatePaddlePredictor(config);
+
+  std::vector<std::vector<PaddleTensor>> input_slots_all;
+  LoadInputData(&input_slots_all);
+
+  std::ifstream fin(FLAGS_refer_result);
+  std::string line;
+  std::vector<float> ref;
+
+  while (std::getline(fin, line)) {
+    Split(line, ' ', &ref);
+  }
+
+  std::vector<PaddleTensor> outputs;
+  for (size_t i = 0; i < input_slots_all.size(); i++) {
+    outputs.clear();
+    predictor->Run(input_slots_all[i], &outputs);
+
+    auto output = outputs.front();
+    size_t outputs_size = 1;
+    for (auto dim : output.shape) {
+      outputs_size *= dim;
+    }
+    float *result = reinterpret_cast<float *>(output.data.data());
+    for (size_t j = 0; j < outputs_size; ++j) {
+      EXPECT_NEAR(ref[i * outputs_size + j], result[j], 5e-2);
+    }
+  }
+}
+
+// Compare results
+TEST(Ernie_gpu_fp16_with_ir, compare_results) {
+  AnalysisConfig config;
+  config.SetModel(FLAGS_infer_model);
+  config.EnableUseGpu(512, 0, paddle_infer::PrecisionType::kHalf);
+  config.SwitchIrOptim(true);
+  // The fc_fuse_pass has diff, which will be repaired later.
+  config.pass_builder()->DeletePass("fc_fuse_pass");
+  // There is a problem with the model itself, which has nothing to do with
+  // constant_folding_pass.
+  config.pass_builder()->DeletePass("constant_folding_pass");
+
+  auto predictor = CreatePaddlePredictor(config);
+
+  std::vector<std::vector<PaddleTensor>> input_slots_all;
+  LoadInputData(&input_slots_all);
+
+  std::ifstream fin(FLAGS_refer_result);
+  std::string line;
+  std::vector<float> ref;
+
+  while (std::getline(fin, line)) {
+    Split(line, ' ', &ref);
+  }
+
+  std::vector<PaddleTensor> outputs;
+  for (size_t i = 0; i < input_slots_all.size(); i++) {
+    outputs.clear();
+    predictor->Run(input_slots_all[i], &outputs);
+
+    auto output = outputs.front();
+    size_t outputs_size = 1;
+    for (auto dim : output.shape) {
+      outputs_size *= dim;
+    }
+    float *result = reinterpret_cast<float *>(output.data.data());
+    for (size_t j = 0; j < outputs_size; ++j) {
+      EXPECT_NEAR(ref[i * outputs_size + j], result[j], 5e-2);
+    }
+  }
+}
+
+// Compare results
+TEST(Ernie_gpu_bf16_no_ir, compare_results) {
+  AnalysisConfig config;
+  config.SetModel(FLAGS_infer_model);
+  config.EnableUseGpu(512, 0, paddle_infer::PrecisionType::kBf16);
+  config.SwitchIrOptim(false);
+
+  auto predictor = CreatePaddlePredictor(config);
+
+  std::vector<std::vector<PaddleTensor>> input_slots_all;
+  LoadInputData(&input_slots_all);
+
+  std::ifstream fin(FLAGS_refer_result);
+  std::string line;
+  std::vector<float> ref;
+
+  while (std::getline(fin, line)) {
+    Split(line, ' ', &ref);
+  }
+
+  std::vector<PaddleTensor> outputs;
+  for (size_t i = 0; i < input_slots_all.size(); i++) {
+    outputs.clear();
+    predictor->Run(input_slots_all[i], &outputs);
+
+    auto output = outputs.front();
+    size_t outputs_size = 1;
+    for (auto dim : output.shape) {
+      outputs_size *= dim;
+    }
+    float *result = reinterpret_cast<float *>(output.data.data());
+    for (size_t j = 0; j < outputs_size; ++j) {
+      EXPECT_NEAR(ref[i * outputs_size + j], result[j], 7e-2);
+    }
+  }
+}
+
+// Compare results
+TEST(Ernie_gpu_bf16_with_ir, compare_results) {
+  AnalysisConfig config;
+  config.SetModel(FLAGS_infer_model);
+  config.EnableUseGpu(512, 0, paddle_infer::PrecisionType::kBf16);
+  config.SwitchIrOptim(true);
+  // The fc_fuse_pass has diff, which will be repaired later.
+  config.pass_builder()->DeletePass("fc_fuse_pass");
+  // There is a problem with the model itself, which has nothing to do with
+  // constant_folding_pass.
+  config.pass_builder()->DeletePass("constant_folding_pass");
+
+  auto predictor = CreatePaddlePredictor(config);
+
+  std::vector<std::vector<PaddleTensor>> input_slots_all;
+  LoadInputData(&input_slots_all);
+
+  std::ifstream fin(FLAGS_refer_result);
+  std::string line;
+  std::vector<float> ref;
+
+  while (std::getline(fin, line)) {
+    Split(line, ' ', &ref);
+  }
+
+  std::vector<PaddleTensor> outputs;
+  for (size_t i = 0; i < input_slots_all.size(); i++) {
+    outputs.clear();
+    predictor->Run(input_slots_all[i], &outputs);
+
+    auto output = outputs.front();
+    size_t outputs_size = 1;
+    for (auto dim : output.shape) {
+      outputs_size *= dim;
+    }
+    float *result = reinterpret_cast<float *>(output.data.data());
+    for (size_t j = 0; j < outputs_size; ++j) {
+      EXPECT_NEAR(ref[i * outputs_size + j], result[j], 7e-2);
+    }
+  }
+}
+
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/tests/api/paddle_infer_api_test.cc

-/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+/* Copyright (c) 2022 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.
@@ -12,15 +12,11 @@ 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. */
 
-#include <cuda_runtime.h>
 #include <glog/logging.h>
 #include <gtest/gtest.h>
 
-#include <cstring>
-#include <numeric>
-
 #include "gflags/gflags.h"
-#include "paddle/fluid/inference/tests/api/trt_test_helper.h"
+#include "paddle/fluid/inference/tests/api/tester_helper.h"
 
 namespace paddle_infer {
 

paddle/fluid/pybind/CMakeLists.txt

@@ -262,10 +262,6 @@ if(WITH_PYTHON)
     list(APPEND OP_FUNCTION_GENERETOR_DEPS cncl_context)
   endif()
 
-  if(NOT ((NOT WITH_PYTHON) AND ON_INFER))
-    list(APPEND OP_FUNCTION_GENERETOR_DEPS ${PYTHON_LIBRARIES})
-  endif()
-
   add_executable(op_function_generator op_function_generator.cc)
   target_link_libraries(op_function_generator ${OP_FUNCTION_GENERETOR_DEPS})
   add_executable(eager_legacy_op_function_generator
@@ -605,13 +601,4 @@ if(WITH_PYTHON)
   get_property(os_dependency_modules GLOBAL PROPERTY OS_DEPENDENCY_MODULES)
   target_link_libraries(${SHARD_LIB_NAME} ${os_dependency_modules})
   add_dependencies(${SHARD_LIB_NAME} op_function_generator_cmd)
-
-  if(APPLE)
-    string(REGEX REPLACE ".+/(.+)" "\\1" PYTHON_LIBRARY_NAME
-                         ${PYTHON_LIBRARIES})
-    # target_link_libraries(${SHARD_LIB_NAME} "-Wl,-rpath,${PYTHON_LIBRARY_NAME}")
-  else()
-    target_link_libraries(${SHARD_LIB_NAME} ${PYTHON_LIBRARIES})
-  endif()
-
 endif()

paddle/fluid/pybind/inference_api.cc

@@ -642,7 +642,8 @@ void BindAnalysisConfig(py::module *m) {
       .def("enable_use_gpu",
            &AnalysisConfig::EnableUseGpu,
            py::arg("memory_pool_init_size_mb"),
-           py::arg("device_id") = 0)
+           py::arg("device_id") = 0,
+           py::arg("precision_mode") = AnalysisConfig::Precision::kFloat32)
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
       .def("set_exec_stream",
            [](AnalysisConfig &self, phi::CUDAStream &stream) {