Expose graph ModelPath to TensorRT shared library (#6353)

* Update graph_viewer.cc

* Update tensorrt_execution_provider.cc

* Update graph_viewer.h

* Update tensorrt_execution_provider.cc

* Update tensorrt_execution_provider.cc

* Update provider_api.h

* Update provider_bridge_ort.cc

* Update provider_interfaces.h

* Update provider_interfaces.h

* expose GraphViewer ModelPath API to TRT shared lib

* add modelpath to compile

* update

* add model_path to onnx tensorrt parser

* use GenerateMetaDefId to generate unique TRT kernel name

* use GenerateMetaDefId to generate unique TRT engine name

* fix issue

* Update tensorrt_execution_provider.cc

* remove GetVecHash

* Update tensorrt_execution_provider.h

* convert wchar_t to char for tensorrt parser

* update tensorrt parser to include latest changes

* fix issues

* Update tensorrt_execution_provider.cc

* merge trt parser latest change

* add PROVIDER_DISALLOW_ALL(Path)

.gitmodules

@@ -56,9 +56,10 @@
 [submodule "cmake/external/SafeInt/safeint"]
 	path = cmake/external/SafeInt/safeint
 	url = https://github.com/dcleblanc/SafeInt.git
-[submodule "cmake/external/onnx-tensorrt"]
-	path = cmake/external/onnx-tensorrt
-	url = https://github.com/onnx/onnx-tensorrt.git
 [submodule "cmake/external/optional-lite"]
 	path = cmake/external/optional-lite
 	url = https://github.com/martinmoene/optional-lite.git
+[submodule "cmake/external/onnx-tensorrt"]
+	path = cmake/external/onnx-tensorrt
+	url = https://github.com/onnx/onnx-tensorrt.git
+	branch = 7.1

cmake/external/onnx-tensorrt

@@ -1 +1 @@
-Subproject commit a3a4e38b2dfa7a62b6dcae33c0d1678b3bb5ef2a
+Subproject commit b3eda616d3bb60dcd40a142e4a0a2ad95a7aa166

onnxruntime/core/framework/provider_bridge_ort.cc

@@ -514,6 +514,7 @@ struct ProviderHostImpl : ProviderHost {
   }
 
   const std::string& GraphViewer__Name(const GraphViewer* p) noexcept override { return p->Name(); }
+  const Path& GraphViewer__ModelPath(const GraphViewer* p) noexcept override { return p->ModelPath(); }
 
   const Node* GraphViewer__GetNode(const GraphViewer* p, NodeIndex node_index) override { return p->GetNode(node_index); }
   const NodeArg* GraphViewer__GetNodeArg(const GraphViewer* p, const std::string& name) override { return p->GetNodeArg(name); }
@@ -535,6 +536,10 @@ struct ProviderHostImpl : ProviderHost {
   const std::vector<NodeIndex>& GraphViewer__GetNodesInTopologicalOrder(const GraphViewer* p) override { return p->GetNodesInTopologicalOrder(); }
   const std::vector<const NodeArg*>& GraphViewer__GetInputsIncludingInitializers(const GraphViewer* p) noexcept override { return p->GetInputsIncludingInitializers(); }
 
+  //Path
+  //Gets a string representation of the path.
+  PathString Path__ToPathString(const Path* p) noexcept override { return p->ToPathString(); }
+
   // Provider_OpKernel_Base
   const OpKernelInfo& Provider_OpKernel_Base__GetInfo(const Provider_OpKernel_Base* p) override { return reinterpret_cast<const OpKernel*>(p)->Info(); }
 

onnxruntime/core/providers/shared_library/provider_api.h

@@ -21,6 +21,7 @@
 #include "core/framework/allocatormgr.h"
 #include "core/framework/tensor_shape.h"
 #include "core/providers/providers.h"
+#include "core/common/path_string.h"
 
 namespace onnxruntime {
 namespace logging {
@@ -73,6 +74,7 @@ struct Function;
 struct Graph;
 struct GraphViewer;
 struct Model;
+struct Path;
 struct Node;
 struct NodeArg;
 struct NodeAttributes;

onnxruntime/core/providers/shared_library/provider_interfaces.h

@@ -22,7 +22,6 @@ using DataType = const std::string*;
 }  // namespace ONNX_NAMESPACE
 
 namespace onnxruntime {
-
 // These types don't directly map to internal types
 struct Provider_KernelCreateInfo;
 struct Provider_OpKernel_Base;
@@ -436,6 +435,7 @@ struct ProviderHost {
   virtual std::unique_ptr<Model> GraphViewer__CreateModel(const GraphViewer* p, const logging::Logger& logger) = 0;
 
   virtual const std::string& GraphViewer__Name(const GraphViewer* p) noexcept = 0;
+  virtual const Path& GraphViewer__ModelPath(const GraphViewer* p) noexcept = 0;
 
   virtual const Node* GraphViewer__GetNode(const GraphViewer* p, NodeIndex node_index) = 0;
   virtual const NodeArg* GraphViewer__GetNodeArg(const GraphViewer* p, const std::string& name) = 0;
@@ -456,6 +456,9 @@ struct ProviderHost {
   virtual const std::vector<NodeIndex>& GraphViewer__GetNodesInTopologicalOrder(const GraphViewer* p) = 0;
   virtual const std::vector<const NodeArg*>& GraphViewer__GetInputsIncludingInitializers(const GraphViewer* p) noexcept = 0;
 
+  // Path
+  virtual PathString Path__ToPathString(const Path* p) noexcept = 0;
+
   // Provider_OpKernel_Base
   virtual const OpKernelInfo& Provider_OpKernel_Base__GetInfo(const Provider_OpKernel_Base* p) = 0;
 
@@ -974,6 +977,7 @@ struct GraphViewer {
   std::unique_ptr<Model> CreateModel(const logging::Logger& logger) const { return g_host->GraphViewer__CreateModel(this, logger); }
 
   const std::string& Name() const noexcept { return g_host->GraphViewer__Name(this); }
+  const Path& ModelPath() const noexcept { return g_host->GraphViewer__ModelPath(this); }
 
   const Node* GetNode(NodeIndex node_index) const { return g_host->GraphViewer__GetNode(this, node_index); }
   const NodeArg* GetNodeArg(const std::string& name) const { return g_host->GraphViewer__GetNodeArg(this, name); }
@@ -1000,6 +1004,13 @@ struct GraphViewer {
   GraphViewer(const GraphViewer&) = delete;
   void operator=(const GraphViewer&) = delete;
 };
+
+struct Path {
+  PathString ToPathString() const noexcept { return g_host->Path__ToPathString(this); }
+
+  PROVIDER_DISALLOW_ALL(Path)
+};
+
 #endif
 
 struct Provider_OpKernel_Base {

onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc

@@ -41,14 +41,6 @@ std::string GetCachePath(const std::string& root, const std::string& name) {
   }
 }
 
-std::string GetVecHash(const std::string& vec) {
-  std::size_t ret = vec.size();
-  for (auto i : vec) {
-    ret ^= static_cast<int>(i) + 0x9e3779b9 + (ret << 6) + (ret >> 2);
-  }
-  return std::to_string(ret);
-}
-
 float ConvertSinglePrecisionIEEE754ToFloat(unsigned long input) {
   int s = (input >> 31) & 0x01;
   int e = ((input & 0x7f800000) >> 23) - 127;
@@ -259,7 +251,6 @@ bool SetDynamicRange(nvinfer1::INetworkDefinition& network, std::unordered_map<s
   }
   return true;
 }
-
 }  // namespace
 
 namespace google {
@@ -380,7 +371,7 @@ TensorrtLogger& GetTensorrtLogger() {
 }
 
 TensorrtExecutionProvider::TensorrtExecutionProvider(const TensorrtExecutionProviderInfo& info)
-    : IExecutionProvider{onnxruntime::kTensorrtExecutionProvider}, device_id_(info.device_id) {
+    : IExecutionProvider{onnxruntime::kTensorrtExecutionProvider, true}, device_id_(info.device_id) {
   CUDA_CALL_THROW(cudaSetDevice(device_id_));
 
   AllocatorCreationInfo default_memory_info(
@@ -600,12 +591,15 @@ std::unique_ptr<IndexedSubGraph> TensorrtExecutionProvider::GetSubGraph(SubGraph
     outputs.insert(std::pair<int, const NodeArg*>(it->second, it->first));
   }
 
-  // Assign inputs and outputs to subgraph's meta_def
+  // Generate unique kernel name for TRT subgraph
+  uint64_t model_hash = 0;
+  int id = GenerateMetaDefId(graph, model_hash);
+  std::string subgraph_id = std::to_string(model_hash) + "_" + std::to_string(id);
   auto meta_def = IndexedSubGraph_MetaDef::Create();
   const std::string graph_type = graph.IsSubgraph() ? "subgraph" : "graph";
-  meta_def->name() = "TRTKernel_" + graph_type + "_" + graph.Name() + "_" + std::to_string(subgraph_id_++);
-  meta_def->domain() = kMSDomain;
+  meta_def->name() = "TRTKernel_" + graph_type + "_" + graph.Name() + "_" + subgraph_id;
 
+  // Assign inputs and outputs to subgraph's meta_def
   for (const auto& input : inputs) {
     if (input.second->Exists()) {
       meta_def->inputs().push_back(input.second->Name());
@@ -618,6 +612,7 @@ std::unique_ptr<IndexedSubGraph> TensorrtExecutionProvider::GetSubGraph(SubGraph
     }
   }
 
+  meta_def->domain() = kMSDomain;
   meta_def->since_version() = 1;
   sub_graph->SetMetaDef(std::move(meta_def));
 
@@ -742,7 +737,7 @@ SubGraphCollection_t TensorrtExecutionProvider::GetSupportedList(SubGraphCollect
         auto trt_network = tensorrt_ptr::unique_pointer<nvinfer1::INetworkDefinition>(trt_builder->createNetworkV2(explicitBatch));
 
         auto trt_parser = tensorrt_ptr::unique_pointer<nvonnxparser::IParser>(nvonnxparser::createParser(*trt_network, trt_logger));
-        trt_parser->supportsModel(string_buf.data(), string_buf.size(), parser_nodes_list);
+        trt_parser->supportsModel(string_buf.data(), string_buf.size(), parser_nodes_list, model_path_);
 
         SubGraphCollection_t next_nodes_list;
         const std::vector<NodeIndex>& subgraph_node_index = graph_viewer->GetNodesInTopologicalOrder();
@@ -875,6 +870,14 @@ void TensorrtExecutionProvider::RemoveTensorRTGraphCycles(SubGraphCollection_t&
 std::vector<std::unique_ptr<ComputeCapability>>
 TensorrtExecutionProvider::GetCapability(const GraphViewer& graph,
                                          const std::vector<const KernelRegistry*>& /*kernel_registries*/) const {
+  // Get ModelPath
+  const auto& path_string = graph.ModelPath().ToPathString();
+#ifdef _WIN32
+  wcstombs(model_path_, path_string.c_str(), sizeof(model_path_));
+#else
+  strcpy(model_path_, path_string.c_str());
+#endif
+
   // Get supported node list from TensorRT parser
   const int number_of_ort_nodes = graph.NumberOfNodes();
   std::vector<size_t> nodes_vector(number_of_ort_nodes);
@@ -948,7 +951,6 @@ common::Status TensorrtExecutionProvider::Compile(const std::vector<Node*>& fuse
     auto graph_body_viewer = graph_body.CreateGraphViewer();
     auto model = graph_body_viewer->CreateModel(*GetLogger());
     auto model_proto = model->ToProto();
-
     *model_proto->mutable_graph() = *graph_body.ToGraphProto();
     model_proto->set_ir_version(ONNX_NAMESPACE::Version::IR_VERSION);
     std::string string_buf;
@@ -966,7 +968,7 @@ common::Status TensorrtExecutionProvider::Compile(const std::vector<Node*>& fuse
     auto trt_network = tensorrt_ptr::unique_pointer<nvinfer1::INetworkDefinition>(trt_builder->createNetworkV2(explicitBatch));
     auto trt_config = tensorrt_ptr::unique_pointer<nvinfer1::IBuilderConfig>(trt_builder->createBuilderConfig());
     auto trt_parser = tensorrt_ptr::unique_pointer<nvonnxparser::IParser>(nvonnxparser::createParser(*trt_network, trt_logger));
-    trt_parser->parse(string_buf.data(), string_buf.size());
+    trt_parser->parse(string_buf.data(), string_buf.size(), model_path_);
     trt_config->setMaxWorkspaceSize(max_workspace_size_);
 
     int num_inputs = trt_network->getNbInputs();
@@ -1037,8 +1039,6 @@ common::Status TensorrtExecutionProvider::Compile(const std::vector<Node*>& fuse
       trt_node_name_with_precision += "_int8";
       LOGS_DEFAULT(VERBOSE) << "[TensorRT EP] INT8 mode is enabled";
     }
-    int num_nodes = graph_body_viewer->NumberOfNodes();
-    trt_node_name_with_precision += "_" + GetVecHash(trt_node_name_with_precision + std::to_string(num_nodes));
 
     // Build TRT engine here if the graph doesn't have dynamic shape input. Otherwise engine will
     // be built at runtime

onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.h

@@ -129,7 +129,7 @@ class TensorrtExecutionProvider : public IExecutionProvider {
   OrtMutex tensorrt_mu_;
   int device_id_;
   AllocatorPtr allocator_;
-  mutable int subgraph_id_ = 0;
+  mutable char model_path_[4096]; // Reserved for max path length
 
   std::unordered_map<std::string, tensorrt_ptr::unique_pointer<nvonnxparser::IParser>> parsers_;
   std::unordered_map<std::string, tensorrt_ptr::unique_pointer<nvinfer1::ICudaEngine>> engines_;