[TensorRT EP] Add stream sync after enqueue (#18026)

If the model is partitioned into TRT subgraphs and CUDA EP node, we
observed cuda stream synchronization issue when multithreading. Calling
stream sync API after enqueue can solve this issue without adding much
performance overhead.

onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc

@@ -1869,6 +1869,7 @@ TensorrtExecutionProvider::GetCapability(const GraphViewer& graph,
   } else if (number_of_trt_nodes == number_of_ort_nodes) {
     LOGS_DEFAULT(INFO) << "[TensorRT EP] Whole graph will run on TensorRT execution provider";
   } else {
+    sync_stream_after_enqueue_ = true;
     LOGS_DEFAULT(INFO) << "[TensorRT EP] Graph is partitioned and number of subgraphs running on TensorRT execution provider is " << number_of_subgraphs;
   }
 
@@ -2387,7 +2388,7 @@ common::Status TensorrtExecutionProvider::Compile(const std::vector<FusedNodeAnd
       *p = {context->allocate_func, context->release_func, context->allocator_handle, context->node_name,
             &parsers_[context->node_name], &engines_[context->node_name], &contexts_[context->node_name], &builders_[context->node_name],
             &networks_[context->node_name], input_info_[context->node_name], output_info_[context->node_name],
-            input_shape_ranges_[context->node_name], &tensorrt_mu_, fp16_enable_, int8_enable_, int8_calibration_cache_available_,
+            input_shape_ranges_[context->node_name], sync_stream_after_enqueue_, &tensorrt_mu_, fp16_enable_, int8_enable_, int8_calibration_cache_available_,
             dla_enable_, dla_core_, &max_workspace_size_, trt_node_name_with_precision, engine_cache_enable_, cache_path_,
             runtime_.get(), profiles_[context->node_name], context_memory_sharing_enable_, &max_ctx_mem_size_,
             dynamic_range_map, engine_decryption_enable_, engine_decryption_, engine_encryption_, timing_cache_enable_,
@@ -2415,6 +2416,7 @@ common::Status TensorrtExecutionProvider::Compile(const std::vector<FusedNodeAnd
       const std::unordered_map<std::string, size_t>& input_indexes = (trt_state->input_info)[0];
       const std::unordered_map<std::string, size_t>& output_indexes = (trt_state->output_info)[0];
       const std::unordered_map<std::string, size_t>& output_types = (trt_state->output_info)[1];
+      bool sync_stream_after_enqueue = trt_state->sync_stream_after_enqueue;
       auto fused_node_name = trt_state->fused_node_name;
       auto& shape_ranges = trt_state->input_shape_ranges;
       auto trt_builder = trt_state->builder->get();
@@ -3022,6 +3024,10 @@ common::Status TensorrtExecutionProvider::Compile(const std::vector<FusedNodeAnd
         return ORT_MAKE_STATUS(ONNXRUNTIME, FAIL, "TensorRT EP execution context enqueue failed.");
       }
 
+      if (sync_stream_after_enqueue) {
+        cudaStreamSynchronize(stream);
+      }
+
       // Cast INT64 input to INT32 because TensorRT doesn't fully support INT64
       for (size_t i = 0, end = output_binding_names.size(); i < end; ++i) {
         const std::string& output_name = output_binding_names[i];

onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.h

@@ -111,6 +111,7 @@ struct TensorrtFuncState {
   std::vector<std::unordered_map<std::string, size_t>> input_info;
   std::vector<std::unordered_map<std::string, size_t>> output_info;
   std::unordered_map<std::string, std::unordered_map<size_t, std::vector<std::vector<int64_t>>>> input_shape_ranges;
+  bool sync_stream_after_enqueue = false;
   OrtMutex* tensorrt_mu_ptr = nullptr;
   bool fp16_enable = false;
   bool int8_enable = false;
@@ -262,6 +263,9 @@ class TensorrtExecutionProvider : public IExecutionProvider {
   cudnnHandle_t external_cudnn_handle_ = nullptr;
   cublasHandle_t external_cublas_handle_ = nullptr;
 
+  // Call cudaStreamSynchronize() after TRT enqueueV2()/enqueueV3()
+  mutable bool sync_stream_after_enqueue_ = false;
+
   CUDAGraph cuda_graph_;
   bool is_graph_captured_ = false;
   int regular_run_count_before_graph_capture_ = 0;