[VitisAI] 1. api compatbile 2. dynamic load onnx (#18470)

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

1. Add a backward-compatible API for compiling model.
2. Run-time load vitisai-ep.dll


### 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: Yueqing Zhang <[email protected]>
Co-authored-by: Zhenze Wang <[email protected]>

cmake/onnxruntime_providers_vitisai.cmake

@@ -15,16 +15,10 @@
     "${ONNXRUNTIME_ROOT}/core/providers/vitisai/imp/*.cc"
     "${ONNXRUNTIME_ROOT}/core/providers/vitisai/imp/*.h"
   )
-  list(REMOVE_ITEM onnxruntime_providers_vitisai_cc_srcs "${ONNXRUNTIME_ROOT}/core/providers/vitisai/onnxruntime_vitisai_ep_stub.cc")
   source_group(TREE ${ONNXRUNTIME_ROOT}/core FILES ${onnxruntime_providers_vitisai_cc_srcs})
   onnxruntime_add_static_library(onnxruntime_providers_vitisai ${onnxruntime_providers_vitisai_cc_srcs})
   onnxruntime_add_include_to_target(onnxruntime_providers_vitisai onnxruntime_common onnxruntime_framework onnx onnx_proto)
-  onnxruntime_add_shared_library(onnxruntime_vitisai_ep ${ONNXRUNTIME_ROOT}/core/providers/vitisai/onnxruntime_vitisai_ep_stub.cc)
-  onnxruntime_add_include_to_target(onnxruntime_vitisai_ep onnxruntime_common)
-  target_include_directories(onnxruntime_vitisai_ep PRIVATE "${ONNXRUNTIME_ROOT}" "${ONNXRUNTIME_ROOT}/core/providers/vitisai/include")
-  target_link_libraries(onnxruntime_providers_vitisai PUBLIC onnxruntime_vitisai_ep PRIVATE onnx protobuf::libprotobuf nlohmann_json::nlohmann_json )
-  target_compile_definitions(onnxruntime_vitisai_ep
-                           PRIVATE "-DONNXRUNTIME_VITISAI_EP_STUB=1" "-DONNXRUNTIME_VITISAI_EP_EXPORT_DLL=1")
+  target_link_libraries(onnxruntime_providers_vitisai PRIVATE onnx protobuf::libprotobuf nlohmann_json::nlohmann_json)
   if(NOT MSVC)
     target_compile_options(onnxruntime_providers_vitisai PUBLIC $<$<CONFIG:DEBUG>:-U_FORTIFY_SOURCE -D_FORTIFY_SOURCE=0>)
   endif(NOT MSVC)
@@ -49,4 +43,4 @@
             LIBRARY   DESTINATION ${CMAKE_INSTALL_LIBDIR}
             RUNTIME   DESTINATION ${CMAKE_INSTALL_BINDIR}
             FRAMEWORK DESTINATION ${CMAKE_INSTALL_BINDIR})
-  endif()
+  endif()

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

@@ -2,6 +2,16 @@
 // Copyright (c) 2023 Advanced Micro Devices, Inc. All rights reserved.
 // Licensed under the MIT License.
 #pragma once
+#include <vector>
+#include <memory>
+#include <string>
+
 #include "core/session/onnxruntime_cxx_api.h"
+#include "core/framework/provider_options.h"
+#include "vaip/my_ort.h"
+#include "vaip/dll_safe.h"
+#include "vaip/custom_op.h"
 
 std::vector<OrtCustomOpDomain*> initialize_vitisai_ep();
+vaip_core::DllSafe<std::vector<std::unique_ptr<vaip_core::ExecutionProvider>>> compile_onnx_model_with_options(
+    const std::string& model_path, const onnxruntime::Graph& graph, const onnxruntime::ProviderOptions& options);

onnxruntime/core/providers/vitisai/vitisai_execution_provider.cc

@@ -15,30 +15,28 @@
 #include "core/session/custom_ops.h"
 #include "core/session/inference_session.h"
 
-#include "onnxruntime_vitisai_ep/onnxruntime_vitisai_ep.h"
-
 using namespace ONNX_NAMESPACE;
 
 namespace onnxruntime {
 
 constexpr const char* VITISAI = "VITISAI";
 
 static vaip_core::DllSafe<std::vector<std::unique_ptr<vaip_core::ExecutionProvider>>> compile_onnx_model(
-    const onnxruntime::GraphViewer& graph_viewer,
-    const logging::Logger& logger, const char* json_config) {
+    const onnxruntime::GraphViewer& graph_viewer, const logging::Logger& logger, const ProviderOptions& options) {
 #ifndef _WIN32
   auto model_path = graph_viewer.ModelPath().ToPathString();
 #else
   using convert_t = std::codecvt_utf8<wchar_t>;
   std::wstring_convert<convert_t, wchar_t> strconverter;
   auto model_path = strconverter.to_bytes(graph_viewer.ModelPath().ToPathString());
 #endif
-  return onnxruntime_vitisai_ep::compile_onnx_model_3(model_path, graph_viewer.GetGraph(), json_config);
+  return compile_onnx_model_with_options(model_path, graph_viewer.GetGraph(), options);
 }
+
 struct MyCustomOpKernel : OpKernel {
   MyCustomOpKernel(const OpKernelInfo& info, const OrtCustomOp& op) : OpKernel(info), op_(op) {
-    op_kernel_ = op_.CreateKernel(&op_, OrtGetApiBase()->GetApi(op_.version),
-                                  reinterpret_cast<const OrtKernelInfo*>(&info));
+    op_kernel_ =
+        op_.CreateKernel(&op_, OrtGetApiBase()->GetApi(op_.version), reinterpret_cast<const OrtKernelInfo*>(&info));
   }
 
   ~MyCustomOpKernel() override { op_.KernelDestroy(op_kernel_); }
@@ -55,8 +53,7 @@ struct MyCustomOpKernel : OpKernel {
   void* op_kernel_;
 };
 
-VitisAIExecutionProvider::VitisAIExecutionProvider(
-    const VitisAIExecutionProviderInfo& info)
+VitisAIExecutionProvider::VitisAIExecutionProvider(const ProviderOptions& info)
     : IExecutionProvider{onnxruntime::kVitisAIExecutionProvider}, info_(info) {
   custom_op_domains_ = initialize_vitisai_ep();
   registry_ = std::make_shared<KernelRegistry>();
@@ -77,7 +74,8 @@ void VitisAIExecutionProvider::CreateKernelRegistry() {
         }
       }
       def_builder.Provider(onnxruntime::kVitisAIExecutionProvider);
-      KernelCreateFn kernel_create_fn = [op](FuncManager&, const OpKernelInfo& info, std::unique_ptr<OpKernel>& out) -> Status {
+      KernelCreateFn kernel_create_fn = [op](FuncManager&, const OpKernelInfo& info,
+                                             std::unique_ptr<OpKernel>& out) -> Status {
         out = std::make_unique<MyCustomOpKernel>(info, *op);
         return Status::OK();
       };
@@ -89,9 +87,8 @@ void VitisAIExecutionProvider::CreateKernelRegistry() {
 
 std::shared_ptr<KernelRegistry> VitisAIExecutionProvider::GetKernelRegistry() const { return registry_; }
 
-std::vector<std::unique_ptr<ComputeCapability>>
-VitisAIExecutionProvider::GetCapability(const onnxruntime::GraphViewer& graph,
-                                        const IKernelLookup& /*kernel_lookup*/) const {
+std::vector<std::unique_ptr<ComputeCapability>> VitisAIExecutionProvider::GetCapability(
+    const onnxruntime::GraphViewer& graph, const IKernelLookup& /*kernel_lookup*/) const {
   if (graph.IsSubgraph()) {
     // VITIS AI EP not support sungraph. Assigned to CPU.
     return {};
@@ -100,9 +97,7 @@ VitisAIExecutionProvider::GetCapability(const onnxruntime::GraphViewer& graph,
     // Only compiling a model once is currently supported
     return {};
   }
-  auto opt_str = info_.get_json_config_str();  // String
-  execution_providers_ =
-      std::make_unique<my_ep_t>(compile_onnx_model(graph, *GetLogger(), opt_str));
+  execution_providers_ = std::make_unique<my_ep_t>(compile_onnx_model(graph, *GetLogger(), info_));
   auto result = vaip::GetComputeCapabilityOps(graph, execution_providers_.get(), vitisai_optypes_);
   size_t index = 0u;
   for (auto& ep : **execution_providers_) {
@@ -112,32 +107,26 @@ VitisAIExecutionProvider::GetCapability(const onnxruntime::GraphViewer& graph,
   return result;
 }
 
-common::Status VitisAIExecutionProvider::Compile(
-    const std::vector<FusedNodeAndGraph>& fused_nodes_and_graphs,
-    std::vector<NodeComputeInfo>& node_compute_funcs) {
+common::Status VitisAIExecutionProvider::Compile(const std::vector<FusedNodeAndGraph>& fused_nodes_and_graphs,
+                                                 std::vector<NodeComputeInfo>& node_compute_funcs) {
   for (const auto& fused_node_graph : fused_nodes_and_graphs) {
     NodeComputeInfo compute_info;
     const onnx::AttributeProto* attr = graph_utils::GetNodeAttribute(fused_node_graph.fused_node, "index");
     assert(attr != nullptr);
     size_t index = (size_t)attr->i();
-    compute_info.create_state_func = [this, index](ComputeContext* context,
-                                                   FunctionState* state) {
+    compute_info.create_state_func = [this, index](ComputeContext* context, FunctionState* state) {
       auto* p = (**this->execution_providers_)[index]->compile().release();
       *state = p;
       return 0;
     };
 
     compute_info.release_state_func = [](FunctionState state) {
       if (state) {
-        delete reinterpret_cast<vaip_core::CustomOp*>(
-            state);
+        delete reinterpret_cast<vaip_core::CustomOp*>(state);
       }
     };
-    compute_info.compute_func = [](FunctionState state, const OrtApi* api,
-                                   OrtKernelContext* context) {
-      reinterpret_cast<vaip_core::CustomOp*>(
-          state)
-          ->Compute(api, context);
+    compute_info.compute_func = [](FunctionState state, const OrtApi* api, OrtKernelContext* context) {
+      reinterpret_cast<vaip_core::CustomOp*>(state)->Compute(api, context);
       return Status::OK();
     };
     node_compute_funcs.push_back(compute_info);

onnxruntime/core/providers/vitisai/vitisai_execution_provider.h

@@ -4,6 +4,10 @@
 #pragma once
 
 #include <ctime>
+#include <vector>
+#include <memory>
+#include <set>
+#include <string>
 
 #include "core/framework/execution_provider.h"
 #include "core/framework/customregistry.h"
@@ -18,34 +22,19 @@ class ExecutionProvider;
 }  // namespace vaip_core
 namespace onnxruntime {
 
-// Information needed to construct execution providers.
-struct VitisAIExecutionProviderInfo {
-  VitisAIExecutionProviderInfo(const ProviderOptions& provider_options);
-
-  const char* get_json_config_str() const {
-    return json_config_.c_str();
-  }
-
- private:
-  ProviderOptions provider_options_;
-  const std::string json_config_;
-};
-
 // Logical device representation.
 class VitisAIExecutionProvider : public IExecutionProvider {
  public:
-  explicit VitisAIExecutionProvider(const VitisAIExecutionProviderInfo& info);
+  explicit VitisAIExecutionProvider(const ProviderOptions& info);
   ~VitisAIExecutionProvider() = default;
 
-  std::vector<std::unique_ptr<ComputeCapability>>
-  GetCapability(const onnxruntime::GraphViewer& graph,
-                const IKernelLookup& /*kernel_lookup*/) const override;
+  std::vector<std::unique_ptr<ComputeCapability>> GetCapability(const onnxruntime::GraphViewer& graph,
+                                                                const IKernelLookup& /*kernel_lookup*/) const override;
 
   int GetDeviceId() const { return 0; }
 
-  common::Status Compile(
-      const std::vector<FusedNodeAndGraph>& fused_nodes_and_graphs,
-      std::vector<NodeComputeInfo>& node_compute_funcs) override;
+  common::Status Compile(const std::vector<FusedNodeAndGraph>& fused_nodes_and_graphs,
+                         std::vector<NodeComputeInfo>& node_compute_funcs) override;
   std::shared_ptr<KernelRegistry> GetKernelRegistry() const override;
 
  private:
@@ -54,7 +43,7 @@ class VitisAIExecutionProvider : public IExecutionProvider {
   using my_ep_uptr_t = std::shared_ptr<my_ep_t>;
   // we have to hide the implementation by forward declaration.
   mutable my_ep_uptr_t execution_providers_;
-  VitisAIExecutionProviderInfo info_;
+  ProviderOptions info_;
   std::vector<OrtCustomOpDomain*> custom_op_domains_;
   std::shared_ptr<KernelRegistry> registry_;
   std::set<std::string> vitisai_optypes_;

onnxruntime/core/providers/vitisai/vitisai_provider_factory.cc

@@ -3,56 +3,37 @@
 
 #include "vitisai_provider_factory_creator.h"
 
+#include <unordered_map>
+#include <string>
+
 #include "vaip/global_api.h"
 #include "./vitisai_execution_provider.h"
 #include "core/framework/execution_provider.h"
 
 #include "core/session/abi_session_options_impl.h"
-#include "nlohmann/json.hpp"
-#include <fstream>
-#include <unordered_map>
-#include <string>
+#include "core/providers/shared_library/provider_host_api.h"
 
 using namespace onnxruntime;
-using json = nlohmann::json;
 namespace onnxruntime {
 
-static std::string ConfigToJsonStr(const std::unordered_map<std::string, std::string>& config) {
-  const auto& filename = config.at("config_file");
-  std::ifstream f(filename);
-  json data = json::parse(f);
-  for (const auto& entry : config) {
-    data[entry.first] = entry.second;
-  }
-  return data.dump();
-}
-
-VitisAIExecutionProviderInfo::VitisAIExecutionProviderInfo(const ProviderOptions& provider_options) : provider_options_(provider_options), json_config_{ConfigToJsonStr(provider_options)} {}
-
 struct VitisAIProviderFactory : IExecutionProviderFactory {
-  VitisAIProviderFactory(const VitisAIExecutionProviderInfo& info) : info_(info) {}
+  VitisAIProviderFactory(const ProviderOptions& info) : info_(info) {}
   ~VitisAIProviderFactory() = default;
 
   std::unique_ptr<IExecutionProvider> CreateProvider() override;
 
  private:
-  VitisAIExecutionProviderInfo info_;
+  ProviderOptions info_;
 };
 
 std::unique_ptr<IExecutionProvider> VitisAIProviderFactory::CreateProvider() {
   return std::make_unique<VitisAIExecutionProvider>(info_);
 }
 
-std::shared_ptr<IExecutionProviderFactory>
-CreateExecutionProviderFactory_VITISAI(const VitisAIExecutionProviderInfo& info) {
-  initialize_vitisai_ep();
-  return std::make_shared<VitisAIProviderFactory>(info);
-}
-
-std::shared_ptr<IExecutionProviderFactory> VitisAIProviderFactoryCreator::Create(const ProviderOptions& provider_options) {
+std::shared_ptr<IExecutionProviderFactory> VitisAIProviderFactoryCreator::Create(
+    const ProviderOptions& provider_options) {
   initialize_vitisai_ep();
-  auto info = VitisAIExecutionProviderInfo{provider_options};
-  return std::make_shared<VitisAIProviderFactory>(info);
+  return std::make_shared<VitisAIProviderFactory>(provider_options);
 }
 
 }  // namespace onnxruntime

onnxruntime/core/providers/vitisai/vitisai_provider_factory_creator.h

@@ -9,9 +9,6 @@
 #include "core/framework/provider_options.h"
 
 namespace onnxruntime {
-
-struct VitisAIExecutionProviderInfo;
-
 struct VitisAIProviderFactoryCreator {
   static std::shared_ptr<IExecutionProviderFactory> Create(const ProviderOptions& provider_options);
 };

onnxruntime/python/onnxruntime_pybind_state_common.h

@@ -85,13 +85,6 @@ struct OrtStatus {
 #define BACKEND_TVM ""
 #endif
 
-#if USE_VITISAI
-#define BACKEND_VITISAI "-VITISAI"
-#include "core/providers/vitisai/vitisai_execution_provider.h"
-#else
-#define BACKEND_VITISAI ""
-#endif
-
 #if USE_OPENBLAS
 #define BACKEND_OPENBLAS "-OPENBLAS"
 #else
@@ -451,9 +444,6 @@ std::shared_ptr<IExecutionProviderFactory> CreateExecutionProviderFactory_Dnnl(c
 std::shared_ptr<IExecutionProviderFactory> CreateExecutionProviderFactory_Tvm(const tvm::TvmEPOptions& info);
 std::shared_ptr<IExecutionProviderFactory> CreateExecutionProviderFactory_Tvm(const char* params);
 #endif
-std::shared_ptr<IExecutionProviderFactory> CreateExecutionProviderFactory_VITISAI(const char* backend_type, int device_id,
-                                                                                  const char* export_runtime_module,
-                                                                                  const char* load_runtime_module);
 std::shared_ptr<IExecutionProviderFactory> CreateExecutionProviderFactory_ACL(int use_arena);
 std::shared_ptr<IExecutionProviderFactory> CreateExecutionProviderFactory_ArmNN(int use_arena);
 std::shared_ptr<IExecutionProviderFactory> CreateExecutionProviderFactory_DML(int device_id);