[VitisAI] Cache node subgraph when necessary (microsoft#22073)

### Description
<!-- Describe your changes. -->

[VitisAI] Cache node subgraph when necessary

### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->

---------

Co-authored-by: Zhenze Wang <zhenzew@xilinx.com>
Co-authored-by: zhenzew <zhenzew@amd.com>

Author: 3 people

cmake/onnxruntime_providers_vitisai.cmake

@@ -12,6 +12,7 @@
   file(GLOB onnxruntime_providers_vitisai_cc_srcs CONFIGURE_DEPENDS
     "${ONNXRUNTIME_ROOT}/core/providers/vitisai/*.cc"
     "${ONNXRUNTIME_ROOT}/core/providers/vitisai/*.h"
+    "${ONNXRUNTIME_ROOT}/core/providers/vitisai/include/vaip/*.h"
     "${ONNXRUNTIME_ROOT}/core/providers/vitisai/imp/*.cc"
     "${ONNXRUNTIME_ROOT}/core/providers/vitisai/imp/*.h"
     "${ONNXRUNTIME_ROOT}/core/providers/shared_library/*.h"

onnxruntime/core/providers/shared_library/provider_interfaces.h

@@ -996,6 +996,7 @@ struct ProviderHost {
                                     bool include_outer_scope_args,
                                     int execution_order) noexcept = 0;
   virtual const Node* GraphViewer__GetProducerNode(const GraphViewer* p, const std::string& node_arg_name) const = 0;
+  virtual IOnnxRuntimeOpSchemaCollectionPtr GraphViewer__GetSchemaRegistry(const GraphViewer* p) const = 0;
 
   // OpKernel
   virtual const Node& OpKernel__Node(const OpKernel* p) = 0;

onnxruntime/core/providers/shared_library/provider_wrappedtypes.h

@@ -1068,6 +1068,7 @@ class GraphViewer final {
     g_host->GraphViewer__ToProto(this, graph_proto, include_initializers, include_outer_scope_args, execution_order);
   }
   const Node* GetProducerNode(const std::string& node_arg_name) const { return g_host->GraphViewer__GetProducerNode(this, node_arg_name); }
+  IOnnxRuntimeOpSchemaCollectionPtr GetSchemaRegistry() const { return g_host->GraphViewer__GetSchemaRegistry(this); }
 
   GraphViewer() = delete;
   GraphViewer(const GraphViewer&) = delete;

onnxruntime/core/providers/vitisai/imp/global_api.cc

@@ -444,6 +444,11 @@ vaip_core::OrtApiForVaip* create_org_api_hook() {
     }
   };
   the_global_api.node_arg_external_location = vaip::node_arg_external_location;
+  the_global_api.model_to_proto = [](onnxruntime::Model& model) { return model.ToProto().release(); };
+  the_global_api.model_proto_serialize_as_string = [](ONNX_NAMESPACE::ModelProto& model_proto) {
+    return vaip_core::DllSafe(model_proto.SerializeAsString());
+  };
+  the_global_api.model_proto_delete = [](ONNX_NAMESPACE::ModelProto* p) { delete p; };
   if (!s_library_vitisaiep.vaip_get_version) {
     return reinterpret_cast<vaip_core::OrtApiForVaip*>(&(the_global_api.host_));
   } else {

onnxruntime/core/providers/vitisai/include/vaip/custom_op.h

@@ -25,18 +25,18 @@ class ExecutionProvider {
   virtual DllSafe<std::vector<std::string>> get_meta_def_nodes() const = 0;
   virtual DllSafe<std::vector<std::string>>
   get_meta_def_constant_initializer() const = 0;
+  virtual bool get_meta_def_fallback_CPU() const { return false; };
   virtual std::unique_ptr<CustomOp> compile() const = 0;
 
  public:
-  inline void set_fused_node(const onnxruntime::Node* fused_node) {
-    fused_node_ = fused_node;
-  }
-  inline const onnxruntime::Node* get_fused_node() const {
-    return fused_node_;
-  }
+  inline void set_fused_node(const onnxruntime::Node* fused_node) { fused_node_ = fused_node; }
+  inline const onnxruntime::Node* get_fused_node() const { return fused_node_; }
+  inline void set_model(onnxruntime::Model* model) { model_ = model; }
+  inline onnxruntime::Model* get_model() const { return model_; }
 
  private:
   const onnxruntime::Node* fused_node_ = nullptr;
+  onnxruntime::Model* model_ = nullptr;
 };
 
 class CustomOp {

onnxruntime/core/providers/vitisai/include/vaip/my_ort.h

@@ -20,6 +20,7 @@ struct NodeAttributes;
 namespace ONNX_NAMESPACE {
 struct AttributeProto;
 struct TensorProto;
+struct ModelProto;
 #ifndef USE_VITISAI
 enum TensorProto_DataType : int {
   TensorProto_DataType_UNDEFINED = 0,
@@ -70,6 +71,7 @@ enum AttributeProto_AttributeType : int {
 namespace vaip_core {
 class GraphHolder;
 using ONNX_NAMESPACE::AttributeProto;
+using ONNX_NAMESPACE::ModelProto;
 using ONNX_NAMESPACE::TensorProto;
 using onnxruntime::Graph;
 using onnxruntime::GraphViewer;

onnxruntime/core/providers/vitisai/include/vaip/vaip_ort_api.h

@@ -13,7 +13,7 @@ struct OrtApi;
 
 namespace vaip_core {
 
-#define VAIP_ORT_API_MAJOR (10u)
+#define VAIP_ORT_API_MAJOR (11u)
 #define VAIP_ORT_API_MINOR (0u)
 #define VAIP_ORT_API_PATCH (0u)
 struct OrtApiForVaip {
@@ -231,6 +231,9 @@ struct OrtApiForVaip {
                            gsl::span<const NodeArg* const> inputs);                                                                                   // [92]
   int (*node_arg_external_location)(const Graph& graph, const NodeArg& node_arg, std::string& file, size_t& offset, size_t& size, size_t& checksum);  // [93]
   void (*session_option_configuration)(void* mmap, void* session_option, void (*push)(void* mmap, const char* name, const char* value));              // [94]
+  ModelProto* (*model_to_proto)(Model& model);                                                                                                        // [95]
+  DllSafe<std::string> (*model_proto_serialize_as_string)(ModelProto& model_proto);                                                                   // [96]
+  void (*model_proto_delete)(ModelProto* p);                                                                                                          // [97]
 };
 
 #ifndef USE_VITISAI

onnxruntime/core/providers/vitisai/vitisai_execution_provider.cc

@@ -76,7 +76,17 @@ common::Status VitisAIExecutionProvider::Compile(const std::vector<FusedNodeAndG
     auto& attrs = fused_node_graph.fused_node.get().GetAttributes();
     assert(attrs.count("index"));
     size_t index = attrs.at("index").i();
-    (**this->execution_providers_)[index]->set_fused_node(&fused_node_graph.fused_node.get());
+    auto& ep = (**this->execution_providers_)[index];
+    ep->set_fused_node(&fused_node_graph.fused_node.get());
+    if (ep->get_meta_def_fallback_CPU()) {
+      auto& subgraph = fused_node_graph.filtered_graph.get();
+      auto& logger = logging::LoggingManager::DefaultLogger();
+      auto model_proto = subgraph.CreateModel(logger)->ToProto();
+      subgraph.ToProto(*model_proto->mutable_graph(), true, true);
+      auto local_registries = IOnnxRuntimeOpSchemaRegistryList{subgraph.GetSchemaRegistry()};
+      auto model = Model::Create(std::move(*model_proto), subgraph.ModelPath(), &local_registries, logger);
+      ep->set_model(model.release());
+    }
     compute_info.create_state_func = [this, index](ComputeContext* context, FunctionState* state) {
       auto* p = (**this->execution_providers_)[index]->compile().release();
       *state = p;

onnxruntime/core/providers/vitisai/vitisai_execution_provider.h

@@ -50,7 +50,6 @@ class VitisAIExecutionProvider : public IExecutionProvider {
   ProviderOptions info_;
   std::vector<OrtCustomOpDomain*> custom_op_domains_;
   std::shared_ptr<KernelRegistry> registry_;
-  std::set<std::string> vitisai_optypes_;
   // EP context related.
   bool ep_ctx_enabled_ = false;
   bool ep_ctx_embed_mode_ = true;

onnxruntime/core/session/provider_bridge_ort.cc

@@ -1212,6 +1212,7 @@ struct ProviderHostImpl : ProviderHost {
     GraphViewerToProto(*p, graph_proto, include_initializers, include_outer_scope_args, static_cast<ExecutionOrder>(execution_order));
   }
   const Node* GraphViewer__GetProducerNode(const GraphViewer* p, const std::string& node_arg_name) const override { return p->GetProducerNode(node_arg_name); }
+  IOnnxRuntimeOpSchemaCollectionPtr GraphViewer__GetSchemaRegistry(const GraphViewer* p) const override { return p->GetSchemaRegistry(); }
 
   // OpKernel (direct)
   const Node& OpKernel__Node(const OpKernel* p) override { return p->OpKernel::Node(); }