PaddlePaddle
diff --git a/‎paddle/fluid/framework/naive_executor.cc
Copy file name to clipboard
+4-2Lines changed: 4 additions & 2 deletions b/‎paddle/fluid/framework/naive_executor.cc
Copy file name to clipboard
+4-2Lines changed: 4 additions & 2 deletions
diff --git a/‎paddle/fluid/framework/naive_executor.h
Copy file name to clipboard
+2-1Lines changed: 2 additions & 1 deletion b/‎paddle/fluid/framework/naive_executor.h
Copy file name to clipboard
+2-1Lines changed: 2 additions & 1 deletion
diff --git a/‎paddle/fluid/framework/new_executor/interpreter_base_impl.h
Copy file name to clipboard
+4-2Lines changed: 4 additions & 2 deletions b/‎paddle/fluid/framework/new_executor/interpreter_base_impl.h
Copy file name to clipboard
+4-2Lines changed: 4 additions & 2 deletions
diff --git a/‎paddle/fluid/framework/new_executor/interpretercore.cc
Copy file name to clipboard
+11-5Lines changed: 11 additions & 5 deletions b/‎paddle/fluid/framework/new_executor/interpretercore.cc
Copy file name to clipboard
+11-5Lines changed: 11 additions & 5 deletions
diff --git a/‎paddle/fluid/framework/new_executor/interpretercore.h
Copy file name to clipboard
+4-2Lines changed: 4 additions & 2 deletions b/‎paddle/fluid/framework/new_executor/interpretercore.h
Copy file name to clipboard
+4-2Lines changed: 4 additions & 2 deletions
diff --git a/‎paddle/fluid/framework/new_executor/pir_interpreter.cc
Copy file name to clipboard
+16-2Lines changed: 16 additions & 2 deletions b/‎paddle/fluid/framework/new_executor/pir_interpreter.cc
Copy file name to clipboard
+16-2Lines changed: 16 additions & 2 deletions
diff --git a/‎paddle/fluid/framework/new_executor/pir_interpreter.h
Copy file name to clipboard
+4-2Lines changed: 4 additions & 2 deletions b/‎paddle/fluid/framework/new_executor/pir_interpreter.h
Copy file name to clipboard
+4-2Lines changed: 4 additions & 2 deletions
diff --git a/‎paddle/fluid/framework/new_executor/program_interpreter.cc
Copy file name to clipboard
+55-41Lines changed: 55 additions & 41 deletions b/‎paddle/fluid/framework/new_executor/program_interpreter.cc
Copy file name to clipboard
+55-41Lines changed: 55 additions & 41 deletions
diff --git a/‎paddle/fluid/framework/new_executor/program_interpreter.h
Copy file name to clipboard
+8-3Lines changed: 8 additions & 3 deletions b/‎paddle/fluid/framework/new_executor/program_interpreter.h
Copy file name to clipboard
+8-3Lines changed: 8 additions & 3 deletions
diff --git a/‎paddle/fluid/inference/api/analysis_predictor.cc
Copy file name to clipboard
+5-5Lines changed: 5 additions & 5 deletions b/‎paddle/fluid/inference/api/analysis_predictor.cc
Copy file name to clipboard
+5-5Lines changed: 5 additions & 5 deletions
@@ -72,12 +72,14 @@ void NaiveExecutor::PrepareInterpreterCore(
 }
 
 void NaiveExecutor::RunInterpreterCore(
-    const std::vector<std::string> &feed_names, bool need_fetch) {
+    const std::vector<std::string> &feed_names,
+    bool need_fetch,
+    bool switch_stream) {
   platform::ScopedFlushDenormal flush;
 #ifdef PADDLE_WITH_NVTX
   platform::CudaNvtxRangePush("model", platform::NvtxRangeColor::Yellow);
 #endif
-  interpreter_core_->Run(feed_names, need_fetch);
+  interpreter_core_->Run(feed_names, need_fetch, false, false, switch_stream);
 #ifdef PADDLE_WITH_NVTX
   platform::CudaNvtxRangePop();
 #endif
 
@@ -77,7 +77,8 @@ class NaiveExecutor {
   void Run();
 
   void RunInterpreterCore(const std::vector<std::string>& feed_names = {},
-                          bool need_fetch = false);
+                          bool need_fetch = false,
+                          bool switch_stream = false);
 
   // Get an tensor to operating directly, without the need for feed_ops.
   phi::DenseTensor* FindTensor(const std::string& name);
 
@@ -68,13 +68,15 @@ class InterpreterBaseImpl {
       const std::vector<std::string>& feed_names,
       const std::vector<phi::DenseTensor>& feed_tensors,
       bool need_fetch = true,
-      bool enable_job_schedule_profiler = false) = 0;
+      bool enable_job_schedule_profiler = false,
+      bool switch_stream = false) = 0;
 
   virtual paddle::framework::FetchList Run(
       const std::vector<std::string>& feed_names,
       bool need_fetch = true,
       bool enable_job_schedule_profiler = false,
-      bool enable_op_profiling = false) = 0;
+      bool enable_op_profiling = false,
+      bool switch_stream = false) = 0;
 
   virtual void ShareWorkQueueFrom(InterpreterBaseImpl* src) = 0;
 
 
@@ -67,19 +67,25 @@ FetchList InterpreterCore::Run(
     const std::vector<std::string>& feed_names,
     const std::vector<phi::DenseTensor>& feed_tensors,
     bool need_fetch,
-    bool enable_job_schedule_profiler) {
-  return impl_->Run(
-      feed_names, feed_tensors, need_fetch, enable_job_schedule_profiler);
+    bool enable_job_schedule_profiler,
+    bool switch_stream) {
+  return impl_->Run(feed_names,
+                    feed_tensors,
+                    need_fetch,
+                    enable_job_schedule_profiler,
+                    switch_stream);
 }
 
 FetchList InterpreterCore::Run(const std::vector<std::string>& feed_names,
                                bool need_fetch,
                                bool enable_job_schedule_profiler,
-                               bool enable_op_profiling) {
+                               bool enable_op_profiling,
+                               bool switch_stream) {
   return impl_->Run(feed_names,
                     need_fetch,
                     enable_job_schedule_profiler,
-                    enable_op_profiling);
+                    enable_op_profiling,
+                    switch_stream);
 }
 
 void InterpreterCore::ShareWorkQueueFrom(std::shared_ptr<InterpreterCore> src) {
 
@@ -49,12 +49,14 @@ class InterpreterCore {
       const std::vector<std::string>& feed_names,
       const std::vector<phi::DenseTensor>& feed_tensors,
       bool need_fetch = true,
-      bool enable_job_schedule_profiler = false);
+      bool enable_job_schedule_profiler = false,
+      bool switch_stream = false);
 
   paddle::framework::FetchList Run(const std::vector<std::string>& feed_names,
                                    bool need_fetch = true,
                                    bool enable_job_schedule_profiler = false,
-                                   bool enable_op_profiling = false);
+                                   bool enable_op_profiling = false,
+                                   bool switch_stream = false);
 
   void RunProfile(const std::vector<std::string>& feed_names);
 
 
@@ -1255,7 +1255,8 @@ paddle::framework::FetchList PirInterpreter::Run(
     const std::vector<std::string>& feed_names,
     const std::vector<phi::DenseTensor>& feed_tensors,
     bool need_fetch,
-    bool enable_job_schedule_profiler) {
+    bool enable_job_schedule_profiler,
+    bool switch_stream) {
   enable_job_schedule_profiler_ = enable_job_schedule_profiler;
 
   auto FeedInput = [&] {
@@ -1318,6 +1319,12 @@ paddle::framework::FetchList PirInterpreter::Run(
     is_build_ = true;
     is_shared_results_build_ = true;
   } else {
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+    if (switch_stream) {
+      BuildInstruction();
+      VLOG(4) << "Done BuildInstruction";
+    }
+#endif
     if (FLAGS_enable_pir_in_executor_trace_run || nccl_op_num_ > 1 ||
         execution_config_.used_for_inference ||
         ((execution_config_.used_for_jit || execution_config_.used_for_cinn) &&
@@ -1350,7 +1357,8 @@ paddle::framework::FetchList PirInterpreter::Run(
 FetchList PirInterpreter::Run(const std::vector<std::string>& feed_names,
                               bool need_fetch,
                               bool enable_job_schedule_profiler,
-                              bool enable_op_profiling) {
+                              bool enable_op_profiling,
+                              bool switch_stream) {
   enable_job_schedule_profiler_ = enable_job_schedule_profiler;
 
   if (enable_op_profiling) {
@@ -1401,6 +1409,12 @@ FetchList PirInterpreter::Run(const std::vector<std::string>& feed_names,
     is_build_ = true;
     is_shared_results_build_ = true;
   } else {
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+    if (switch_stream) {
+      BuildInstruction();
+      VLOG(4) << "Done BuildInstruction";
+    }
+#endif
     if (FLAGS_enable_pir_in_executor_trace_run || nccl_op_num_ > 1 ||
         execution_config_.used_for_inference ||
         ((execution_config_.used_for_jit || execution_config_.used_for_cinn) &&
 
@@ -57,12 +57,14 @@ class PirInterpreter : public InterpreterBaseImpl {
       const std::vector<std::string>& feed_names,
       const std::vector<phi::DenseTensor>& feed_tensors,
       bool need_fetch = true,
-      bool enable_job_schedule_profiler = false) override;
+      bool enable_job_schedule_profiler = false,
+      bool switch_stream = false) override;
 
   paddle::framework::FetchList Run(const std::vector<std::string>& feed_names,
                                    bool need_fetch = true,
                                    bool enable_job_schedule_profiler = false,
-                                   bool enable_op_profiling = false) override;
+                                   bool enable_op_profiling = false,
+                                   bool switch_stream = false) override;
 
   void ShareWorkQueueFrom(InterpreterBaseImpl* src) override;
 
 
@@ -144,7 +144,8 @@ void ProgramInterpreter::RunImpl() {
 FetchList ProgramInterpreter::Run(const std::vector<std::string>& feed_names,
                                   bool need_fetch,
                                   bool enable_job_schedule_profiler,
-                                  bool enable_op_profiling) {
+                                  bool enable_op_profiling,
+                                  bool switch_stream) {
   enable_job_schedule_profiler_ = enable_job_schedule_profiler;
   is_in_op_profiling_mode_ = enable_op_profiling;
 
@@ -163,6 +164,11 @@ FetchList ProgramInterpreter::Run(const std::vector<std::string>& feed_names,
     is_build_ = true;
     is_shared_results_build_ = true;
   } else {
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+    if (switch_stream) {
+      BuildOpFuncNode(&op_func_nodes);
+    }
+#endif
     RunImpl();
   }
 
@@ -233,7 +239,8 @@ FetchList ProgramInterpreter::Run(
     const std::vector<std::string>& feed_names,
     const std::vector<phi::DenseTensor>& feed_tensors,
     bool need_fetch,
-    bool enable_job_schedule_profiler) {
+    bool enable_job_schedule_profiler,
+    bool switch_stream) {
   enable_job_schedule_profiler_ = enable_job_schedule_profiler;
 
   SetDeviceId(place_);
@@ -244,7 +251,7 @@ FetchList ProgramInterpreter::Run(
 #endif
 
   bool is_build = is_build_;
-  Prepare(feed_names, feed_tensors, is_build);
+  Prepare(feed_names, feed_tensors, is_build, switch_stream);
 
   if (is_build) {
     RunImpl();
@@ -671,42 +678,7 @@ std::tuple<double, double> ProgramInterpreter::InterpreterRunTime() {
 void ProgramInterpreter::Convert(
     std::vector<paddle::framework::OpFuncNode>* op_func_nodes) {
   auto& vec_meta_info = var_scope_.MutableVecMetaInfo();
-  auto nodes = *op_func_nodes;
-  auto op_nums = nodes.size();
-  vec_instruction_.clear();
-  vec_instruction_.reserve(op_nums);
-  for (size_t op_idx = 0; op_idx < op_nums; ++op_idx) {
-    auto& op_func_node = nodes[op_idx];
-    stream_analyzer_.SetForceEventsToWaitInfo(force_evnets_to_wait_);
-    auto* dev_ctx_ = stream_analyzer_.ParseDeviceContext(op_func_node);
-#ifdef PADDLE_WITH_CUDA
-    if (FLAGS_new_executor_use_cuda_graph) {
-      auto& op = op_func_node.operator_base_;
-      auto& op_type = op->Type();
-      if (op_type == interpreter::kMemcpyD2H ||
-          op_type == interpreter::kMemcpyH2D) {
-        PADDLE_THROW(paddle::platform::errors::Fatal(
-            "Cuda memory copy d2h/h2d is not allowed while using cuda graph."));
-      }
-      PADDLE_ENFORCE_EQ(typeid(*dev_ctx_) == typeid(phi::GPUContext),
-                        true,
-                        platform::errors::InvalidArgument(
-                            "Device context of op %s must be [%s] while using "
-                            "cuda graph, but got [%s].",
-                            op_type,
-                            typeid(phi::GPUContext).name(),
-                            typeid(*dev_ctx_).name()));
-      // cuda graph needs to record all stream
-      phi::backends::gpu::CUDAGraphContextManager::Instance()
-          .RecordCapturingDeviceContext(dev_ctx_);
-    }
-#endif
-    vec_instruction_.emplace_back(op_idx, std::move(op_func_node), *dev_ctx_);
-
-#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
-    vec_instruction_.back().UpdataRecordStreamForGcInfo();
-#endif
-  }
+  BuildOpFuncNode(op_func_nodes);
 
   BuildOperatorDependences();
 
@@ -743,6 +715,7 @@ void ProgramInterpreter::Convert(
   }
 
   // calculate last_live_ops_
+  auto op_nums = (*op_func_nodes).size();
   for (size_t op_idx = 0; op_idx < op_nums; ++op_idx) {
     Instruction& instr = vec_instruction_[op_idx];
     OpInOutInfo info;
@@ -879,6 +852,46 @@ void ProgramInterpreter::Convert(
   AnalyseExecuteOrderForTrace();
 }
 
+void ProgramInterpreter::BuildOpFuncNode(
+    std::vector<paddle::framework::OpFuncNode>* op_func_nodes) {
+  auto nodes = *op_func_nodes;
+  auto op_nums = nodes.size();
+  vec_instruction_.clear();
+  vec_instruction_.reserve(op_nums);
+  for (size_t op_idx = 0; op_idx < op_nums; ++op_idx) {
+    auto& op_func_node = nodes[op_idx];
+    stream_analyzer_.SetForceEventsToWaitInfo(force_evnets_to_wait_);
+    auto* dev_ctx_ = stream_analyzer_.ParseDeviceContext(op_func_node);
+#ifdef PADDLE_WITH_CUDA
+    if (FLAGS_new_executor_use_cuda_graph) {
+      auto& op = op_func_node.operator_base_;
+      auto& op_type = op->Type();
+      if (op_type == interpreter::kMemcpyD2H ||
+          op_type == interpreter::kMemcpyH2D) {
+        PADDLE_THROW(paddle::platform::errors::Fatal(
+            "Cuda memory copy d2h/h2d is not allowed while using cuda graph."));
+      }
+      PADDLE_ENFORCE_EQ(typeid(*dev_ctx_) == typeid(phi::GPUContext),
+                        true,
+                        platform::errors::InvalidArgument(
+                            "Device context of op %s must be [%s] while using "
+                            "cuda graph, but got [%s].",
+                            op_type,
+                            typeid(phi::GPUContext).name(),
+                            typeid(*dev_ctx_).name()));
+      // cuda graph needs to record all stream
+      phi::backends::gpu::CUDAGraphContextManager::Instance()
+          .RecordCapturingDeviceContext(dev_ctx_);
+    }
+#endif
+    vec_instruction_.emplace_back(op_idx, std::move(op_func_node), *dev_ctx_);
+
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+    vec_instruction_.back().UpdataRecordStreamForGcInfo();
+#endif
+  }
+}
+
 void ProgramInterpreter::BuildSkipShareLoDInfo() {
   for (size_t i = 0; i < vec_instruction_.size(); ++i) {
     bool can_skip_lod = true;
@@ -1494,7 +1507,8 @@ void ProgramInterpreter::CheckGC(const Instruction& instr) {
 void ProgramInterpreter::Prepare(
     const std::vector<std::string>& feed_names,
     const std::vector<phi::DenseTensor>& feed_tensors,
-    bool prepare_feed) {
+    bool prepare_feed,
+    bool switch_stream) {
   PADDLE_ENFORCE_EQ(feed_names.size(),
                     feed_tensors.size(),
                     platform::errors::PreconditionNotMet(
@@ -1517,7 +1531,7 @@ void ProgramInterpreter::Prepare(
     }
   };
 
-  if (!is_build_) {
+  if (!is_build_ || switch_stream) {
     paddle::framework::interpreter::BuildVariableScope(
         block_, execution_config_, &var_scope_);
     FeedInput();
 
@@ -49,12 +49,14 @@ class ProgramInterpreter : public InterpreterBaseImpl {
       const std::vector<std::string>& feed_names,
       const std::vector<phi::DenseTensor>& feed_tensors,
       bool need_fetch = true,
-      bool enable_job_schedule_profiler = false) override;
+      bool enable_job_schedule_profiler = false,
+      bool switch_stream = false) override;
 
   paddle::framework::FetchList Run(const std::vector<std::string>& feed_names,
                                    bool need_fetch = true,
                                    bool enable_job_schedule_profiler = false,
-                                   bool enable_op_profiling = false) override;
+                                   bool enable_op_profiling = false,
+                                   bool switch_stream = false) override;
 
   std::shared_ptr<ProgramDesc> GetMutableCopyProgram() override;
 
@@ -125,6 +127,8 @@ class ProgramInterpreter : public InterpreterBaseImpl {
   void BuildSkipShareLoDInfo();
   void UpdateSyncOpNum();
   void AnalyseExecuteOrderForTrace();
+  void BuildOpFuncNode(
+      std::vector<paddle::framework::OpFuncNode>* op_func_nodes);
 
   // inplace
   void BuildInplace();
@@ -150,7 +154,8 @@ class ProgramInterpreter : public InterpreterBaseImpl {
   // only used when program contains no feed op
   void Prepare(const std::vector<std::string>& feed_names,
                const std::vector<phi::DenseTensor>& feed_tensors,
-               bool prepare_feed);
+               bool prepare_feed,
+               bool switch_stream = false);
 
   void RecordMemcpyD2H(const Instruction& instr_node);
 
 
@@ -2249,7 +2249,7 @@ std::unique_ptr<ZeroCopyTensor> AnalysisPredictor::GetOutputTensor(
   return res;
 }
 
-bool AnalysisPredictor::ZeroCopyRun() {
+bool AnalysisPredictor::ZeroCopyRun(bool switch_stream) {
   inference::DisplayMemoryInfo(place_, "before run");
 #if defined(PADDLE_WITH_DISTRIBUTE) && defined(PADDLE_WITH_PSCORE)
   if (config_.dist_config().use_dist_model()) {
@@ -2312,7 +2312,7 @@ bool AnalysisPredictor::ZeroCopyRun() {
 #endif
 
   if (config_.new_executor_enabled()) {
-    executor_->RunInterpreterCore();
+    executor_->RunInterpreterCore({}, false, switch_stream);
   } else {
     executor_->Run();
   }
@@ -2353,7 +2353,7 @@ bool AnalysisPredictor::ExpRunWithExternalStream(const gpuStream_t stream) {
         "Please use config.SetExecStream to init gpu resources, and then we "
         "will bind gpu resources to execution stream."));
   }
-
+  bool switch_stream = false;
   if (stream != predictor_stream_) {
 #ifdef PADDLE_WITH_HIP
     hipStreamSynchronize(static_cast<gpuStream_t>(predictor_stream_));
@@ -2383,9 +2383,9 @@ bool AnalysisPredictor::ExpRunWithExternalStream(const gpuStream_t stream) {
         }));
     auto &pool = paddle::experimental::DeviceContextPool::Instance();
     pool.SyncDeviceContext(place_);
+    switch_stream = true;
   }
-
-  return ZeroCopyRun();
+  return ZeroCopyRun(switch_stream);
 }
 #endif