Fix compile errors/issues

Author: chilo-ms

plugin_execution_providers/tensorrt/onnx_ctx_model_helper.cc

@@ -5,7 +5,7 @@
 #include <fstream>
 #include <filesystem>
 
-#include "tensorrt_execution_provider_utils.h"
+#include "ep_utils.h"
 #include "onnx_ctx_model_helper.h"
 
 extern TensorrtLogger& GetTensorrtLogger(bool verbose_log);
@@ -109,3 +109,20 @@ OrtStatus* EPContextNodeHelper::CreateEPContextNode(const std::string& engine_ca
 
   return nullptr;
 }
+
+/*
+ * Get the weight-refitted engine cache path from a weight-stripped engine cache path
+ *
+ * Weight-stipped engine:
+ * An engine with weights stripped and its size is smaller than a regualr engine.
+ * The cache name of weight-stripped engine is TensorrtExecutionProvider_TRTKernel_XXXXX.stripped.engine
+ *
+ * Weight-refitted engine:
+ * An engine that its weights have been refitted and it's simply a regular engine.
+ * The cache name of weight-refitted engine is TensorrtExecutionProvider_TRTKernel_XXXXX.engine
+ */
+std::string GetWeightRefittedEnginePath(std::string stripped_engine_cache) {
+  std::filesystem::path stripped_engine_cache_path(stripped_engine_cache);
+  std::string refitted_engine_cache_path = stripped_engine_cache_path.stem().stem().string() + ".engine";
+  return refitted_engine_cache_path;
+}

plugin_execution_providers/tensorrt/onnx_ctx_model_helper.h

@@ -3,8 +3,8 @@
 
 #pragma once
 
-#include "tensorrt_execution_provider_utils.h"
 #include "tensorrt_execution_provider.h"
+#include "ep_utils.h"
 #include "nv_includes.h"
 
 #include <string>

plugin_execution_providers/tensorrt/tensorrt_execution_provider.cc

@@ -12,12 +12,13 @@
 #define ORT_EP_UTILS_ORT_GRAPH_TO_PROTO_IMPL
 #include "ort_graph_to_proto.h"
 
-//#include "tensorrt_execution_provider_utils.h"
+#include "tensorrt_execution_provider_utils.h"
 #include "tensorrt_execution_provider.h"
 #include "cuda_allocator.h"
 #include "onnx_ctx_model_helper.h"
 #include "onnx/onnx_pb.h"
 #include "cuda/unary_elementwise_ops_impl.h"
+#include "ep_utils.h"
 
 #ifdef _WIN32
 #include <windows.h>
@@ -31,6 +32,10 @@
 #define LIBFUNC(lib, fn) dlsym((lib), (fn))
 #endif
 
+const OrtApi* g_ort_api = nullptr;
+const OrtEpApi* g_ep_api = nullptr;
+const OrtModelEditorApi* g_model_editor_api = nullptr;
+
 void CUDA_RETURN_IF_ERROR(cudaError_t res) {
   if (res != cudaSuccess) abort();
 }
@@ -1795,9 +1800,9 @@ OrtStatus* ORT_API_CALL TensorrtExecutionProvider::CompileImpl(_In_ OrtEp* this_
 
     OrtStatus* status;
     if (EPContextNodeHelper::GraphHasCtxNode(graphs[fused_node_idx], ort_api)) {
-      RETURN_IF_ERROR(ep->CreateNodeComputeInfoFromPrecompiledEngine(this_ptr, graphs[fused_node_idx], fused_node,
-                                                                     input_map, output_map,
-                                                                     &node_compute_infos_result[fused_node_idx]));
+      //RETURN_IF_ERROR(ep->CreateNodeComputeInfoFromPrecompiledEngine(this_ptr, graphs[fused_node_idx], fused_node,
+      //                                                               input_map, output_map,
+      //                                                               &node_compute_infos_result[fused_node_idx]));
     } else {
       RETURN_IF_ERROR(ep->CreateNodeComputeInfoFromGraph(this_ptr, graphs[fused_node_idx], fused_node, input_map,
                                                          output_map, &node_compute_infos_result[fused_node_idx],
@@ -1899,6 +1904,8 @@ OrtStatus* TensorrtExecutionProvider::RefitEngine(
 #endif
 }
 
+TensorrtExecutionProvider::~TensorrtExecutionProvider() = default;
+
 /// <summary>
 /// 
 /// Plugin TensorRT EP that implements OrtEp
@@ -1908,7 +1915,8 @@ TensorrtExecutionProvider::TensorrtExecutionProvider(TensorrtExecutionProviderFa
                                                      const std::string& name,
                                                      const OrtSessionOptions& session_options,
                                                      const OrtLogger& logger)
-    : ApiPtrs{static_cast<const ApiPtrs&>(factory)},
+    : OrtEp{},  // explicitly call the struct ctor to ensure all optional values are default initialized
+      ApiPtrs{static_cast<const ApiPtrs&>(factory)},
       factory_(factory),
       name_{name},
       session_options_{session_options},

plugin_execution_providers/tensorrt/tensorrt_execution_provider.def

@@ -0,0 +1,5 @@
+LIBRARY "TensorRTEp.dll"
+EXPORTS
+ CreateEpFactories @1
+ ReleaseEpFactory @2
+ 

plugin_execution_providers/tensorrt/tensorrt_execution_provider.lds

@@ -0,0 +1,7 @@
+VERS_1.0.0 {
+ global:
+  CreateEpFactories;
+  ReleaseEpFactory;
+ local:
+    *;
+};

plugin_execution_providers/tensorrt/tensorrt_execution_provider_data_transfer.h

@@ -3,7 +3,7 @@
 
 #pragma once
 
-#include "tensorrt_execution_provider_utils.h"
+#include "ep_utils.h"
 
 struct TRTEpDataTransfer : OrtDataTransferImpl, ApiPtrs {
   TRTEpDataTransfer(ApiPtrs api_ptrs, std::vector<const OrtMemoryDevice*> device_mem_infos,

plugin_execution_providers/tensorrt/tensorrt_execution_provider_info.cc

@@ -6,6 +6,7 @@
 #include "tensorrt_execution_provider_info.h"
 #include "provider_options_utils.h"
 #include "cuda/cuda_common.h"
+#include "ep_utils.h"
 
 namespace tensorrt {
 namespace provider_option_names {

plugin_execution_providers/tensorrt/tensorrt_execution_provider_info.h

@@ -3,7 +3,6 @@
 
 #pragma once
 
-#include "tensorrt_execution_provider_utils.h"
 #include "provider_options.h"
 
 #include <string>

plugin_execution_providers/tensorrt/tensorrt_execution_provider_utils.h

@@ -4,6 +4,7 @@
 #include "onnxruntime_cxx_api.h"
 #undef ORT_API_MANUAL_INIT
 
+#include "ep_utils.h"
 #include "flatbuffers/idl.h"
 #include "ort_trt_int8_cal_table.fbs.h"
 #include "make_string.h"
@@ -22,104 +23,8 @@
 #include <iostream>
 #include <filesystem>
 
-struct ApiPtrs {
-  const OrtApi& ort_api;
-  const OrtEpApi& ep_api;
-  const OrtModelEditorApi& model_editor_api;
-};
-
-const OrtApi* g_ort_api = nullptr;
-const OrtEpApi* g_ep_api = nullptr;
-const OrtModelEditorApi* g_model_editor_api = nullptr;
-
-#define ENFORCE(condition, ...)                            \
-  do {                                                      \
-    if (!(condition)) {                                      \
-      throw std::runtime_error(MakeString(__VA_ARGS__));    \
-    }                                                       \
-  } while (false)
-
-#define THROW(...) \
-  throw std::runtime_error(MakeString(__VA_ARGS__));
-
-#define RETURN_IF_ERROR(fn)    \
-  do {                         \
-    OrtStatus* _status = (fn); \
-    if (_status != nullptr) {  \
-      return _status;          \
-    }                          \
-  } while (0)
-
-/*
-template <typename... Args>
-std::string ComposeString(Args&&... args) {
-  std::ostringstream oss;
-  (oss << ... << args);
-  return oss.str();
-};
-*/
-
-#define RETURN_IF(cond, ...)                                                               \
-  do {                                                                                     \
-    if ((cond)) {                                                                          \
-      return Ort::GetApi().CreateStatus(ORT_EP_FAIL, MakeString(__VA_ARGS__).c_str()); \
-    }                                                                                      \
-  } while (0)
-
-#define RETURN_IF_NOT(condition, ...) RETURN_IF(!(condition), __VA_ARGS__)
-
-#define MAKE_STATUS(error_code, msg) \
-  Ort::GetApi().CreateStatus(error_code, (msg));
-
-#define THROW_IF_ERROR(expr)                         \
-  do {                                               \
-    auto _status = (expr);                           \
-    if (_status != nullptr) {                        \
-      std::ostringstream oss;                        \
-      oss << Ort::GetApi().GetErrorMessage(_status); \
-      Ort::GetApi().ReleaseStatus(_status);          \
-      throw std::runtime_error(oss.str());           \
-    }                                                \
-  } while (0)
-
-#define RETURN_FALSE_AND_PRINT_IF_ERROR(fn)                            \
-  do {                                                                 \
-    OrtStatus* status = (fn);                                          \
-    if (status != nullptr) {                                           \
-      std::cerr << Ort::GetApi().GetErrorMessage(status) << std::endl; \
-      return false;                                                    \
-    }                                                                  \
-  } while (0)  
-
-// Helper to release Ort one or more objects obtained from the public C API at the end of their scope.
-template <typename T>
-struct DeferOrtRelease {
-  DeferOrtRelease(T** object_ptr, std::function<void(T*)> release_func)
-      : objects_(object_ptr), count_(1), release_func_(release_func) {}
-
-  DeferOrtRelease(T** objects, size_t count, std::function<void(T*)> release_func)
-      : objects_(objects), count_(count), release_func_(release_func) {}
-
-  ~DeferOrtRelease() {
-    if (objects_ != nullptr && count_ > 0) {
-      for (size_t i = 0; i < count_; ++i) {
-        if (objects_[i] != nullptr) {
-          release_func_(objects_[i]);
-          objects_[i] = nullptr;
-        }
-      }
-    }
-  }
-  T** objects_ = nullptr;
-  size_t count_ = 0;
-  std::function<void(T*)> release_func_ = nullptr;
-};
-
 namespace fs = std::filesystem;
 
-template <typename T>
-using AllocatorUniquePtr = std::unique_ptr<T, std::function<void(T*)>>;
-
 bool CalcMemSizeForArrayWithAlignment(size_t nmemb, size_t size, size_t alignment, size_t* out) noexcept {
   size_t alloc_size = size;
   if (alignment == 0) {

plugin_execution_providers/tensorrt/tensorrt_provider_factory.cc

@@ -58,11 +58,16 @@ OrtStatus* ORT_API_CALL TensorrtExecutionProviderFactory::GetSupportedDevicesImp
 
   std::vector<const OrtMemoryDevice*> cuda_gpu_mem_devices;
   std::vector<const OrtMemoryDevice*> cuda_pinned_mem_devices;
+  int GPU_cnt = 0;
 
   for (size_t i = 0; i < num_devices && num_ep_devices < max_ep_devices; ++i) {
     // C API
     const OrtHardwareDevice& device = *devices[i];
     if (factory->ort_api.HardwareDevice_Type(&device) == OrtHardwareDeviceType::OrtHardwareDeviceType_GPU) {
+      if (GPU_cnt > 0) {
+        continue;
+      }
+      GPU_cnt++;
       // These can be returned as nullptr if you have nothing to add.
       OrtKeyValuePairs* ep_metadata = nullptr;
       OrtKeyValuePairs* ep_options = nullptr;
@@ -87,7 +92,8 @@ OrtStatus* ORT_API_CALL TensorrtExecutionProviderFactory::GetSupportedDevicesImp
       }
 
       uint32_t vendor_id = factory->ort_api.HardwareDevice_VendorId(&device);
-      uint32_t device_id = factory->ort_api.HardwareDevice_DeviceId(&device);
+      //uint32_t device_id = factory->ort_api.HardwareDevice_DeviceId(&device);
+      uint32_t device_id = 0;
 
       // CUDA allocator OrtMemoryInfo
       OrtMemoryInfo* mem_info = nullptr;