[onnxruntime/build] Add new flag enable_generic_interface to build primary EPs by default

cmake/CMakeLists.txt

@@ -259,6 +259,12 @@ option(onnxruntime_USE_AZURE "Build with azure inferencing support" OFF)
 option(onnxruntime_USE_LOCK_FREE_QUEUE "Build with lock-free task queue for threadpool." OFF)
 option(onnxruntime_FORCE_GENERIC_ALGORITHMS "Disable optimized arch-specific algorithms. Use only for testing and debugging generic algorithms." OFF)
 
+option(onnxruntime_USE_TENSORRT_INTERFACE "Build ONNXRuntime shared lib which is compatible with TensorRT EP interface" OFF)
+option(onnxruntime_USE_CUDA_INTERFACE "Build ONNXRuntime shared lib which is compatible with Cuda EP interface" OFF)
+option(onnxruntime_USE_OPENVINO_INTERFACE "Build ONNXRuntime shared lib which is compatible with OpenVINO EP interface" OFF)
+option(onnxruntime_USE_VITISAI_INTERFACE "Build ONNXRuntime shared lib which is compatible with Vitis-AI EP interface" OFF)
+option(onnxruntime_USE_QNN_INTERFACE "Build ONNXRuntime shared lib which is compatible with QNN EP interface" OFF)
+
 # ENABLE_TRAINING includes all training functionality
 # The following 2 entry points
 # 1. ORTModule
@@ -703,7 +709,7 @@ if (WIN32)
     # structure was padded due to __declspec(align())
     list(APPEND ORT_WARNING_FLAGS "/wd4324")
     # warning C4800: Implicit conversion from 'X' to bool. Possible information loss
-    if (onnxruntime_USE_OPENVINO)
+    if (onnxruntime_USE_OPENVINO OR onnxruntime_USE_OPENVINO_INTERFACE)
        list(APPEND ORT_WARNING_FLAGS "/wd4800")
     endif()
     # operator 'operator-name': deprecated between enumerations of different types
@@ -864,7 +870,7 @@ else()
   set(onnxruntime_USE_MEMORY_EFFICIENT_ATTENTION OFF)
 endif()
 
-if (onnxruntime_USE_CUDA)
+if (onnxruntime_USE_CUDA OR onnxruntime_USE_CUDA_INTERFACE)
     list(APPEND ORT_PROVIDER_FLAGS -DUSE_CUDA=1)
     list(APPEND ORT_PROVIDER_CMAKE_FLAGS -Donnxruntime_USE_CUDA=1)
     list(APPEND ONNXRUNTIME_PROVIDER_NAMES cuda)
@@ -888,7 +894,7 @@ if (onnxruntime_USE_CUDA)
     endif()
 endif()
 
-if (onnxruntime_USE_VITISAI)
+if (onnxruntime_USE_VITISAI OR onnxruntime_USE_VITISAI_INTERFACE)
     list(APPEND ORT_PROVIDER_FLAGS -DUSE_VITISAI=1)
     list(APPEND ORT_PROVIDER_CMAKE_FLAGS -Donnxruntime_USE_VITISAI=1)
     list(APPEND ONNXRUNTIME_PROVIDER_NAMES vitisai)
@@ -898,12 +904,12 @@ if (onnxruntime_USE_DNNL)
     list(APPEND ONNXRUNTIME_PROVIDER_NAMES dnnl)
     list(APPEND ORT_PROVIDER_CMAKE_FLAGS -Donnxruntime_USE_DNNL=1)
 endif()
-if (onnxruntime_USE_OPENVINO)
+if (onnxruntime_USE_OPENVINO OR onnxruntime_USE_OPENVINO_INTERFACE)
     list(APPEND ORT_PROVIDER_FLAGS -DUSE_OPENVINO=1)
     list(APPEND ONNXRUNTIME_PROVIDER_NAMES openvino)
     list(APPEND ORT_PROVIDER_CMAKE_FLAGS -Donnxruntime_USE_OPENVINO=1)
 endif()
-if (onnxruntime_USE_TENSORRT)
+if (onnxruntime_USE_TENSORRT OR onnxruntime_USE_TENSORRT_INTERFACE)
     list(APPEND ORT_PROVIDER_FLAGS -DUSE_TENSORRT=1)
     #TODO: remove the following line and change the test code in onnxruntime_shared_lib_test to use the new EP API.
     list(APPEND ONNXRUNTIME_PROVIDER_NAMES tensorrt)
@@ -929,7 +935,7 @@ if (onnxruntime_USE_JSEP)
     list(APPEND ORT_PROVIDER_CMAKE_FLAGS -Donnxruntime_USE_JSEP=1)
     list(APPEND ONNXRUNTIME_PROVIDER_NAMES js)
 endif()
-if (onnxruntime_USE_QNN)
+if (onnxruntime_USE_QNN OR onnxruntime_USE_QNN_INTERFACE)
     list(APPEND ORT_PROVIDER_FLAGS -DUSE_QNN=1)
     list(APPEND ORT_PROVIDER_CMAKE_FLAGS -Donnxruntime_USE_QNN=1)
     list(APPEND ONNXRUNTIME_PROVIDER_NAMES qnn)
@@ -957,7 +963,7 @@ if (onnxruntime_USE_QNN)
       endif()
     endif()
 
-    if (MSVC OR ${CMAKE_SYSTEM_NAME} STREQUAL "Linux")
+    if ((NOT onnxruntime_USE_QNN_INTERFACE) AND (MSVC OR ${CMAKE_SYSTEM_NAME} STREQUAL "Linux"))
       file(GLOB QNN_LIB_FILES LIST_DIRECTORIES false "${onnxruntime_QNN_HOME}/lib/${QNN_ARCH_ABI}/libQnn*.so"
 	                                     "${onnxruntime_QNN_HOME}/lib/${QNN_ARCH_ABI}/Qnn*.dll"
                                              "${onnxruntime_QNN_HOME}/lib/${QNN_ARCH_ABI}/libHtpPrepare.so"
@@ -1416,7 +1422,7 @@ if (onnxruntime_ENABLE_TRAINING_APIS)
   )
 endif()
 
-if (onnxruntime_USE_OPENVINO)
+if (onnxruntime_USE_OPENVINO OR onnxruntime_USE_OPENVINO_INTERFACE)
 
   add_definitions(-DUSE_OPENVINO=1)
 
@@ -1429,7 +1435,7 @@ if (onnxruntime_USE_OPENVINO)
     add_definitions(-DOPENVINO_CONFIG_GPU=1)
   endif()
 
-  if (onnxruntime_USE_OPENVINO_CPU)
+  if (onnxruntime_USE_OPENVINO_CPU OR onnxruntime_USE_OPENVINO_INTERFACE) # OpenVino CPU interface is default built.
     add_definitions(-DOPENVINO_CONFIG_CPU=1)
   endif()
 

onnxruntime/core/providers/shared_library/provider_interfaces.h

@@ -178,7 +178,6 @@ struct ProviderHost {
   virtual std::string demangle(const char* name) = 0;
   virtual std::string demangle(const std::string& name) = 0;
 
-#ifdef USE_CUDA
   virtual std::unique_ptr<IAllocator> CreateCUDAAllocator(int16_t device_id, const char* name) = 0;
   virtual std::unique_ptr<IAllocator> CreateCUDAPinnedAllocator(const char* name) = 0;
   virtual std::unique_ptr<IDataTransfer> CreateGPUDataTransfer() = 0;
@@ -190,7 +189,6 @@ struct ProviderHost {
 
   virtual Status CudaCall_false(int retCode, const char* exprString, const char* libName, int successCode, const char* msg, const char* file, const int line) = 0;
   virtual void CudaCall_true(int retCode, const char* exprString, const char* libName, int successCode, const char* msg, const char* file, const int line) = 0;
-#endif
 
 #ifdef USE_MIGRAPHX
   virtual std::unique_ptr<IAllocator> CreateMIGraphXAllocator(int16_t device_id, const char* name) = 0;
@@ -200,7 +198,6 @@ struct ProviderHost {
 #ifdef USE_ROCM
   virtual std::unique_ptr<IAllocator> CreateROCMAllocator(int16_t device_id, const char* name) = 0;
   virtual std::unique_ptr<IAllocator> CreateROCMPinnedAllocator(const char* name) = 0;
-  virtual std::unique_ptr<IDataTransfer> CreateGPUDataTransfer() = 0;
 
   virtual void rocm__Impl_Cast(void* stream, const int64_t* input_data, int32_t* output_data, size_t count) = 0;
   virtual void rocm__Impl_Cast(void* stream, const int32_t* input_data, int64_t* output_data, size_t count) = 0;
@@ -1256,9 +1253,7 @@ struct ProviderHost {
   virtual training::DistributedRunContext& GetDistributedRunContextInstance() = 0;
 #endif
 
-#if defined(USE_CUDA) || defined(USE_ROCM)
   virtual PhiloxGenerator& PhiloxGenerator__Default() = 0;
-#endif
 
 #ifdef ENABLE_TRAINING_TORCH_INTEROP
   virtual void contrib__PythonOpBase__Init(contrib::PythonOpBase* p, const OpKernelInfo& info) = 0;

onnxruntime/core/session/provider_bridge_ort.cc

@@ -258,10 +258,8 @@ struct ProviderHostImpl : ProviderHost {
   void* CPUAllocator__Alloc(CPUAllocator* p, size_t size) override { return p->CPUAllocator::Alloc(size); }
   void CPUAllocator__Free(CPUAllocator* p, void* allocation) override { return p->CPUAllocator::Free(allocation); }
 
-#ifdef USE_CUDA
   std::unique_ptr<IAllocator> CreateCUDAAllocator(int16_t device_id, const char* name) override { return GetProviderInfo_CUDA().CreateCUDAAllocator(device_id, name); }
   std::unique_ptr<IAllocator> CreateCUDAPinnedAllocator(const char* name) override { return GetProviderInfo_CUDA().CreateCUDAPinnedAllocator(name); }
-  std::unique_ptr<IDataTransfer> CreateGPUDataTransfer() override { return GetProviderInfo_CUDA().CreateGPUDataTransfer(); }
 
   void cuda__Impl_Cast(void* stream, const int64_t* input_data, int32_t* output_data, size_t count) override { return GetProviderInfo_CUDA().cuda__Impl_Cast(stream, input_data, output_data, count); }
   void cuda__Impl_Cast(void* stream, const int32_t* input_data, int64_t* output_data, size_t count) override { return GetProviderInfo_CUDA().cuda__Impl_Cast(stream, input_data, output_data, count); }
@@ -271,7 +269,6 @@ struct ProviderHostImpl : ProviderHost {
 
   Status CudaCall_false(int retCode, const char* exprString, const char* libName, int successCode, const char* msg, const char* file, const int line) override { return GetProviderInfo_CUDA().CudaCall_false(retCode, exprString, libName, successCode, msg, file, line); }
   void CudaCall_true(int retCode, const char* exprString, const char* libName, int successCode, const char* msg, const char* file, const int line) override { GetProviderInfo_CUDA().CudaCall_true(retCode, exprString, libName, successCode, msg, file, line); }
-#endif
 
 #ifdef USE_MIGRAPHX
   std::unique_ptr<IAllocator> CreateMIGraphXAllocator(int16_t device_id, const char* name) override { return GetProviderInfo_MIGraphX().CreateMIGraphXAllocator(device_id, name); }
@@ -291,6 +288,8 @@ struct ProviderHostImpl : ProviderHost {
 
   Status RocmCall_false(int retCode, const char* exprString, const char* libName, int successCode, const char* msg, const char* file, const int line) override { return GetProviderInfo_ROCM().RocmCall_false(retCode, exprString, libName, successCode, msg, file, line); }
   void RocmCall_true(int retCode, const char* exprString, const char* libName, int successCode, const char* msg, const char* file, const int line) override { GetProviderInfo_ROCM().RocmCall_true(retCode, exprString, libName, successCode, msg, file, line); }
+#else
+  std::unique_ptr<IDataTransfer> CreateGPUDataTransfer() override { return GetProviderInfo_CUDA().CreateGPUDataTransfer(); }
 #endif
 
   std::string GetEnvironmentVar(const std::string& var_name) override { return Env::Default().GetEnvironmentVar(var_name); }
@@ -1560,9 +1559,7 @@ struct ProviderHostImpl : ProviderHost {
   training::DistributedRunContext& GetDistributedRunContextInstance() override { return training::DistributedRunContext::GetInstance(); }
 #endif
 
-#if defined(USE_CUDA) || defined(USE_ROCM)
   PhiloxGenerator& PhiloxGenerator__Default() override { return PhiloxGenerator::Default(); }
-#endif
 
 #ifdef ENABLE_TRAINING_TORCH_INTEROP
   void contrib__PythonOpBase__Init(contrib::PythonOpBase* p, const OpKernelInfo& info) override { p->PythonOpBase::Init(info); }

tools/ci_build/build.py

@@ -782,6 +782,12 @@
     parser.add_argument("--use_triton_kernel", action="store_true", help="Use triton compiled kernels")
     parser.add_argument("--use_lock_free_queue", action="store_true", help="Use lock-free task queue for threadpool.")
 
+    parser.add_argument(
+        "--enable_generic_interface",
+        action="store_true",
+        help="build ORT shared library and compatible bridge with primary EPs(tensorRT, OpenVino, Qnn, vitisai) but not tests",
+    )
+
     if not is_windows():
         parser.add_argument(
             "--allow_running_as_root",
@@ -1042,6 +1048,12 @@
         "-Donnxruntime_USE_TENSORRT=" + ("ON" if args.use_tensorrt else "OFF"),
         "-Donnxruntime_USE_TENSORRT_BUILTIN_PARSER="
         + ("ON" if args.use_tensorrt_builtin_parser and not args.use_tensorrt_oss_parser else "OFF"),
+        # interface variables are used only for building onnxruntime/onnxruntime_shared.dll but not EPs
+        "-Donnxruntime_USE_TENSORRT_INTERFACE=" + ("ON" if args.enable_generic_interface else "OFF"),
+        "-Donnxruntime_USE_CUDA_INTERFACE=" + ("ON" if args.enable_generic_interface else "OFF"),
+        "-Donnxruntime_USE_OPENVINO_INTERFACE=" + ("ON" if args.enable_generic_interface else "OFF"),
+        "-Donnxruntime_USE_VITISAI_INTERFACE=" + ("ON" if args.enable_generic_interface else "OFF"),
+        "-Donnxruntime_USE_QNN_INTERFACE=" + ("ON" if args.enable_generic_interface else "OFF"),
         # set vars for migraphx
         "-Donnxruntime_USE_MIGRAPHX=" + ("ON" if args.use_migraphx else "OFF"),
         "-Donnxruntime_DISABLE_CONTRIB_OPS=" + ("ON" if args.disable_contrib_ops else "OFF"),
@@ -1372,6 +1384,8 @@
             cmake_args += ["-Donnxruntime_BUILD_QNN_EP_STATIC_LIB=ON"]
         if args.android and args.use_qnn != "static_lib":
             raise BuildError("Only support Android + QNN builds with QNN EP built as a static library.")
+        if args.use_qnn == "static_lib" and args.enable_generic_interface:
+            raise BuildError("Generic ORT interface only supported with QNN EP built as a shared library.")
 
     if args.use_coreml:
         cmake_args += ["-Donnxruntime_USE_COREML=ON"]
@@ -1529,6 +1543,12 @@
             "-Donnxruntime_USE_FULL_PROTOBUF=ON",
         ]
 
+    # When this flag is enabled, that means we only build ONNXRuntime shared library, expecting some compatible EP
+    # shared lib being build in a seperate process. So we skip the test for now as ONNXRuntime shared lib built under
+    # this flag is not expected to work alone
+    if args.enable_generic_interface:
+        cmake_args += ["-Donnxruntime_BUILD_UNIT_TESTS=OFF"]
+
     if args.enable_lazy_tensor:
         import torch
 
@@ -2649,6 +2669,9 @@
         # Disable ONNX Runtime's builtin memory checker
         args.disable_memleak_checker = True
 
+    if args.enable_generic_interface:
+        args.test = False
+
     # If there was no explicit argument saying what to do, default
     # to update, build and test (for native builds).
     if not (args.update or args.clean or args.build or args.test or args.gen_doc):
@@ -2752,7 +2775,10 @@
     source_dir = os.path.normpath(os.path.join(script_dir, "..", ".."))
 
     # if using cuda, setup cuda paths and env vars
-    cuda_home, cudnn_home = setup_cuda_vars(args)
+    cuda_home = ""
+    cudnn_home = ""
+    if args.use_cuda:
+        cuda_home, cudnn_home = setup_cuda_vars(args)
 
     mpi_home = args.mpi_home
     nccl_home = args.nccl_home
@@ -2765,10 +2791,14 @@
     armnn_home = args.armnn_home
     armnn_libs = args.armnn_libs
 
-    qnn_home = args.qnn_home
+    qnn_home = ""
+    if args.use_qnn:
+        qnn_home = args.qnn_home
 
     # if using tensorrt, setup tensorrt paths
-    tensorrt_home = setup_tensorrt_vars(args)
+    tensorrt_home = ""
+    if args.use_tensorrt:
+        tensorrt_home = setup_tensorrt_vars(args)
 
     # if using migraphx, setup migraphx paths
     migraphx_home = setup_migraphx_vars(args)
@@ -2853,9 +2883,9 @@
                     toolset = "host=" + host_arch + ",version=" + args.msvc_toolset
                 else:
                     toolset = "host=" + host_arch
-                if args.cuda_version:
+                if args.use_cuda and args.cuda_version:
                     toolset += ",cuda=" + args.cuda_version
-                elif args.cuda_home:
+                elif args.use_cuda and args.cuda_home:
                     toolset += ",cuda=" + args.cuda_home
                 if args.windows_sdk_version:
                     target_arch += ",version=" + args.windows_sdk_version

tools/ci_build/github/azure-pipelines/win-ci-pipeline.yml

@@ -177,6 +177,25 @@ stages:
         WITH_CACHE: false
         MachinePool: 'onnxruntime-Win-CPU-2022'
 
+- stage: x64_release_ep_generic_interface
+  dependsOn: []
+  jobs:
+    - template: templates/jobs/win-ci-vs-2022-job.yml
+      parameters:
+        BuildConfig: 'RelWithDebInfo'
+        buildArch: x64
+        additionalBuildFlags: --enable_generic_interface
+        msbuildPlatform: x64
+        isX86: false
+        job_name_suffix: x64_release_ep_generic_interface
+        RunOnnxRuntimeTests: false  # --enable_generic_interface does not build tests
+        EnablePython: false
+        isTraining: false
+        ORT_EP_NAME: CPU
+        GenerateDocumentation: false
+        WITH_CACHE: false
+        MachinePool: 'onnxruntime-Win-CPU-2022'
+
 - stage: x86_release
   dependsOn: []
   jobs: