Incorporate code review comments and simply the changes

by introducing EP specific Interface flags

Author: Karim Vadsariya

cmake/CMakeLists.txt

@@ -634,7 +634,7 @@ 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) #TODO[Karim] applys to all projects...
+    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
@@ -657,7 +657,7 @@ if (WIN32)
             set(CMAKE_${type}_LINKER_FLAGS_MINSIZEREL "${CMAKE_${type}_LINKER_FLAGS_MINSIZEREL} /INCREMENTAL:NO")
           endif()
         endif()
-        if (onnxruntime_ENABLE_LTO AND NOT onnxruntime_USE_CUDA)
+        if (onnxruntime_ENABLE_LTO AND NOT onnxruntime_USE_CUDA) #TODO[Low] Any changes ?
           set(CMAKE_${type}_LINKER_FLAGS_RELEASE "${CMAKE_${type}_LINKER_FLAGS_RELEASE} /LTCG")
           set(CMAKE_${type}_LINKER_FLAGS_RELWITHDEBINFO "${CMAKE_${type}_LINKER_FLAGS_RELWITHDEBINFO} /LTCG")
           set(CMAKE_${type}_LINKER_FLAGS_MINSIZEREL "${CMAKE_${type}_LINKER_FLAGS_MINSIZEREL} /LTCG")
@@ -758,7 +758,7 @@ set(ONNXRUNTIME_PROVIDER_NAMES cpu)
 set(ORT_PROVIDER_FLAGS)
 set(ORT_PROVIDER_CMAKE_FLAGS)
 
-if (onnxruntime_USE_CUDA AND (NOT onnxruntime_ENABLE_GENERIC_INTERFACE))
+if (onnxruntime_USE_CUDA)
   enable_language(CUDA)
   message( STATUS "CMAKE_CUDA_COMPILER_VERSION: ${CMAKE_CUDA_COMPILER_VERSION}")
 
@@ -791,7 +791,7 @@ 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)  #TODO[Karim]  Not used anywhere
     list(APPEND ONNXRUNTIME_PROVIDER_NAMES cuda)
@@ -815,7 +815,7 @@ 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)
@@ -825,12 +825,12 @@ 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)
@@ -856,7 +856,7 @@ 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) #TODO[Karim]
+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)
@@ -884,7 +884,7 @@ if (onnxruntime_USE_QNN) #TODO[Karim]
       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"
@@ -1033,7 +1033,7 @@ function(onnxruntime_set_compile_flags target_name)
 
     set_target_properties(${target_name} PROPERTIES COMPILE_WARNING_AS_ERROR ON)
 
-    if (onnxruntime_USE_CUDA AND (NOT onnxruntime_ENABLE_GENERIC_INTERFACE))
+    if (onnxruntime_USE_CUDA)
       # Suppress a "conversion_function_not_usable" warning in gsl/span
       target_compile_options(${target_name} PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:SHELL:-Xcudafe \"--diag_suppress=conversion_function_not_usable\">")
       target_compile_definitions(${target_name} PRIVATE -DDISABLE_CUSPARSE_DEPRECATED)
@@ -1111,7 +1111,7 @@ function(onnxruntime_set_compile_flags target_name)
     if("${CMAKE_C_COMPILER_ID}" STREQUAL "GNU" AND CMAKE_C_COMPILER_VERSION VERSION_LESS 13 AND CMAKE_C_COMPILER_VERSION VERSION_GREATER_EQUAL 12)
       target_compile_options(${target_name} PRIVATE "$<$<COMPILE_LANGUAGE:CXX>:-Wno-maybe-uninitialized>")
     endif()
-    if (onnxruntime_USE_CUDA AND (NOT onnxruntime_ENABLE_GENERIC_INTERFACE))
+    if (onnxruntime_USE_CUDA)
       foreach(FLAG ${ORT_WARNING_FLAGS})
         target_compile_options(${target_name} PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:SHELL:--compiler-options ${FLAG}>")
       endforeach()
@@ -1333,7 +1333,7 @@ if (onnxruntime_ENABLE_TRAINING_APIS)
   )
 endif()
 
-if (onnxruntime_USE_OPENVINO) #TODO[Karim] applies to both onnxruntime.dll and onnxruntime_shared.dll 
+if (onnxruntime_USE_OPENVINO OR onnxruntime_USE_OPENVINO_INTERFACE) 
 
   add_definitions(-DUSE_OPENVINO=1)
 
@@ -1346,7 +1346,7 @@ if (onnxruntime_USE_OPENVINO) #TODO[Karim] applies to both onnxruntime.dll and o
     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()
 
@@ -1412,7 +1412,7 @@ string(APPEND ORT_BUILD_INFO ", cmake cxx flags: ${CMAKE_CXX_FLAGS}")
 configure_file(onnxruntime_config.h.in ${CMAKE_CURRENT_BINARY_DIR}/onnxruntime_config.h)
 get_property(onnxruntime_GENERATOR_IS_MULTI_CONFIG GLOBAL PROPERTY GENERATOR_IS_MULTI_CONFIG)
 
-if (onnxruntime_USE_CUDA AND (NOT onnxruntime_ENABLE_GENERIC_INTERFACE))
+if (onnxruntime_USE_CUDA)
   set(CMAKE_CUDA_RUNTIME_LIBRARY Shared)
   set(CMAKE_CUDA_STANDARD 17)
   if(onnxruntime_CUDA_HOME)
@@ -1686,7 +1686,7 @@ if (onnxruntime_BUILD_SHARED_LIB OR onnxruntime_BUILD_APPLE_FRAMEWORK)
   list(APPEND ONNXRUNTIME_CMAKE_FILES onnxruntime)
 endif()
 
-if (onnxruntime_BUILD_JAVA) #TODO[Karim]  *cmake files included below uses onnxruntime_USE_CUDA, USE_QNN etc variables, 
+if (onnxruntime_BUILD_JAVA) 
   message(STATUS "Java Build is enabled")
   list(APPEND ONNXRUNTIME_CMAKE_FILES onnxruntime_java)
 endif()

cmake/adjust_global_compile_flags.cmake

@@ -283,7 +283,7 @@ if (MSVC)
     string(APPEND CMAKE_C_FLAGS " /arch:AVX512")
   endif()
 
-  if (onnxruntime_ENABLE_LTO AND NOT onnxruntime_USE_CUDA)
+  if (onnxruntime_ENABLE_LTO AND NOT onnxruntime_USE_CUDA) #TODO[Low] Any changes ?
     set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /Gw /GL")
     set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "${CMAKE_CXX_FLAGS_RELWITHDEBINFO} /Gw /GL")
     set(CMAKE_CXX_FLAGS_MINSIZEREL "${CMAKE_CXX_FLAGS_MINSIZEREL} /Gw /GL")

cmake/external/onnxruntime_external_deps.cmake

@@ -400,7 +400,7 @@ if ((CPUINFO_SUPPORTED OR onnxruntime_USE_XNNPACK) AND NOT ANDROID)
   endif()
 endif()
 
-if(onnxruntime_USE_CUDA) #TODO[Karim]
+if(onnxruntime_USE_CUDA) #TODO[Low] Any changes?
   FetchContent_Declare(
     GSL
     URL ${DEP_URL_microsoft_gsl}
@@ -683,7 +683,7 @@ endif()
 
 set(onnxruntime_LINK_DIRS)
 
-if (onnxruntime_USE_CUDA AND (NOT onnxruntime_ENABLE_GENERIC_INTERFACE))
+if (onnxruntime_USE_CUDA)
   find_package(CUDAToolkit REQUIRED)
 
   if(onnxruntime_CUDNN_HOME)

cmake/onnxruntime.cmake

@@ -208,7 +208,7 @@ set(onnxruntime_INTERNAL_LIBRARIES
   ${PROVIDERS_COREML}
   ${PROVIDERS_DML}
   ${PROVIDERS_NNAPI}
-  ${PROVIDERS_QNN}  #TODO[Karim] ??
+  ${PROVIDERS_QNN}
   ${PROVIDERS_SNPE}
   ${PROVIDERS_RKNPU}
   ${PROVIDERS_VSINPU}

cmake/onnxruntime_providers.cmake

@@ -59,12 +59,6 @@ function(add_op_reduction_include_dirs target)
 endfunction()
 
 
-if(onnxruntime_USE_VITISAI) #TODO[Karim] seems like not used??
-  set(PROVIDERS_VITISAI onnxruntime_providers_vitisai)
-endif()
-if(onnxruntime_USE_CUDA) #TODO[Karim]
-  set(PROVIDERS_CUDA onnxruntime_providers_cuda)
-endif()
 if(onnxruntime_USE_COREML)
   set(PROVIDERS_COREML onnxruntime_providers_coreml coreml_proto)
 endif()
@@ -74,7 +68,7 @@ endif()
 if(onnxruntime_USE_JSEP)
   set(PROVIDERS_JS onnxruntime_providers_js)
 endif()
-if(onnxruntime_USE_QNN) #TODO[Karim]
+if(onnxruntime_USE_QNN OR onnxruntime_USE_QNN_INTERFACE) #TODO[Low] Revisit when qnn EP becomes dynamic lib
   set(PROVIDERS_QNN onnxruntime_providers_qnn)
 endif()
 if(onnxruntime_USE_RKNPU)
@@ -123,23 +117,23 @@ if(onnxruntime_USE_SNPE)
 endif()
 
 include(onnxruntime_providers_cpu.cmake)
-if (onnxruntime_USE_CUDA AND (NOT onnxruntime_ENABLE_GENERIC_INTERFACE))
+if (onnxruntime_USE_CUDA)
   include(onnxruntime_providers_cuda.cmake)
 endif()
 
 if (onnxruntime_USE_DNNL)
   include(onnxruntime_providers_dnnl.cmake)
 endif()
 
-if (onnxruntime_USE_TENSORRT AND (NOT onnxruntime_ENABLE_GENERIC_INTERFACE))
+if (onnxruntime_USE_TENSORRT)
   include(onnxruntime_providers_tensorrt.cmake)
 endif()
 
-if (onnxruntime_USE_VITISAI AND (NOT onnxruntime_ENABLE_GENERIC_INTERFACE))
+if (onnxruntime_USE_VITISAI)
   include(onnxruntime_providers_vitisai.cmake)
 endif()
 
-if (onnxruntime_USE_OPENVINO AND (NOT onnxruntime_ENABLE_GENERIC_INTERFACE))
+if (onnxruntime_USE_OPENVINO)
   include(onnxruntime_providers_openvino.cmake)
 endif()
 
@@ -163,8 +157,7 @@ if (onnxruntime_USE_JSEP)
   include(onnxruntime_providers_js.cmake)
 endif()
 
-if (onnxruntime_USE_QNN)
-  message("Hitting qnn check")
+if (onnxruntime_USE_QNN OR onnxruntime_USE_QNN_INTERFACE) #TODO[Low] Revisit when QNN EP becomes dynamic lib.
   include(onnxruntime_providers_qnn.cmake)
 endif()
 

cmake/onnxruntime_python.cmake

@@ -881,7 +881,7 @@ if (onnxruntime_USE_DNNL)
   )
 endif()
 
-if (onnxruntime_USE_VITISAI AND (NOT onnxruntime_ENABLE_GENERIC_INTERFACE))
+if (onnxruntime_USE_VITISAI)
   add_custom_command(
     TARGET onnxruntime_pybind11_state POST_BUILD
     COMMAND ${CMAKE_COMMAND} -E copy
@@ -891,7 +891,7 @@ if (onnxruntime_USE_VITISAI AND (NOT onnxruntime_ENABLE_GENERIC_INTERFACE))
   )
 endif()
 
-if (onnxruntime_USE_TENSORRT AND (NOT onnxruntime_ENABLE_GENERIC_INTERFACE))
+if (onnxruntime_USE_TENSORRT)
   add_custom_command(
     TARGET onnxruntime_pybind11_state POST_BUILD
     COMMAND ${CMAKE_COMMAND} -E copy
@@ -911,7 +911,7 @@ if (onnxruntime_USE_MIGRAPHX)
   )
 endif()
 
-if (onnxruntime_USE_OPENVINO AND (NOT onnxruntime_ENABLE_GENERIC_INTERFACE))
+if (onnxruntime_USE_OPENVINO)
     add_custom_command(
       TARGET onnxruntime_pybind11_state POST_BUILD
       COMMAND ${CMAKE_COMMAND} -E copy
@@ -934,7 +934,7 @@ if (DEFINED ENV{OPENVINO_MANYLINUX})
     )
 endif()
 
-if (onnxruntime_USE_CUDA AND (NOT onnxruntime_ENABLE_GENERIC_INTERFACE))
+if (onnxruntime_USE_CUDA)
     add_custom_command(
       TARGET onnxruntime_pybind11_state POST_BUILD
       COMMAND ${CMAKE_COMMAND} -E copy

tools/ci_build/build.py

@@ -999,6 +999,8 @@ def generate_build_tree(
     disable_optional_type = "optional" in types_to_disable
     disable_sparse_tensors = "sparsetensor" in types_to_disable
 
+    enable_qnn_interface = True if((args.arm64 or args.arm or args.arm64ec) and (args.enable_generic_interface)) else False
+
     cmake_args += [
         "-Donnxruntime_RUN_ONNX_TESTS=" + ("ON" if args.enable_onnx_tests else "OFF"),
         "-Donnxruntime_GENERATE_TEST_REPORTS=ON",
@@ -1026,7 +1028,11 @@ def generate_build_tree(
         "-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"),
-        "-Donnxruntime_ENABLE_GENERIC_INTERFACE=" + ("ON" if args.enable_generic_interface 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 and not enable_qnn_interface) else "OFF"),   
+        "-Donnxruntime_USE_OPENVINO_INTERFACE=" + ("ON" if (args.enable_generic_interface and not enable_qnn_interface) else "OFF"),
+        "-Donnxruntime_USE_VITISAI_INTERFACE=" + ("ON" if (args.enable_generic_interface and not enable_qnn_interface) else "OFF"),
+        "-Donnxruntime_USE_QNN_INTERFACE=" + ("ON" if (args.enable_generic_interface and enable_qnn_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"),
@@ -1143,8 +1149,6 @@ def generate_build_tree(
         cmake_args.append("-Donnxruntime_ENABLE_WEBASSEMBLY_SIMD=" + ("ON" if args.enable_wasm_simd else "OFF"))
     if args.use_migraphx:
         cmake_args.append("-Donnxruntime_MIGRAPHX_HOME=" + migraphx_home)
-        
-    '''
     if args.use_cuda:
         nvcc_threads = number_of_nvcc_threads(args)
         cmake_args.append("-Donnxruntime_NVCC_THREADS=" + str(nvcc_threads))
@@ -1155,14 +1159,11 @@ def generate_build_tree(
                     f"Add '--disable_types float8' to your command line. See option disable_types."
                 )
         cmake_args.append(f"-DCMAKE_CUDA_COMPILER={cuda_home}/bin/nvcc")
-    '''
-
     if args.use_rocm:
         cmake_args.append("-Donnxruntime_ROCM_HOME=" + rocm_home)
         cmake_args.append("-Donnxruntime_ROCM_VERSION=" + args.rocm_version)
-    
-    #if args.use_tensorrt:
-    #    cmake_args.append("-Donnxruntime_TENSORRT_HOME=" + tensorrt_home)
+    if args.use_tensorrt:
+        cmake_args.append("-Donnxruntime_TENSORRT_HOME=" + tensorrt_home)
 
     if args.use_cuda:
         add_default_definition(cmake_extra_defines, "onnxruntime_USE_CUDA", "ON")
@@ -1214,8 +1215,8 @@ def generate_build_tree(
     if nccl_home and os.path.exists(nccl_home):
         cmake_args += ["-Donnxruntime_NCCL_HOME=" + nccl_home]
 
-    #if qnn_home and os.path.exists(qnn_home):
-    #    cmake_args += ["-Donnxruntime_QNN_HOME=" + qnn_home]
+    if qnn_home and os.path.exists(qnn_home):
+        cmake_args += ["-Donnxruntime_QNN_HOME=" + qnn_home]
 
     if snpe_root and os.path.exists(snpe_root):
         cmake_args += ["-DSNPE_ROOT=" + snpe_root]
@@ -1305,8 +1306,8 @@ def generate_build_tree(
             cmake_args += ["-DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=" + args.xcode_code_signing_team_id]
 
     if args.use_qnn:
-        #if args.qnn_home is None or os.path.exists(args.qnn_home) is False:
-        #    raise BuildError("qnn_home=" + qnn_home + " not valid." + " qnn_home paths must be specified and valid.")
+        if args.qnn_home is None or os.path.exists(args.qnn_home) is False:
+            raise BuildError("qnn_home=" + qnn_home + " not valid." + " qnn_home paths must be specified and valid.")
         cmake_args += ["-Donnxruntime_USE_QNN=ON"]
 
     if args.use_coreml:
@@ -2580,10 +2581,6 @@ def main():
 
     if args.enable_generic_interface:
         args.test = False
-        args.use_tensorrt = True
-        args.use_openvino = "CPU"
-        args.use_vitisai = True
-        #args.use_qnn = True    #defaults should be set based on arm64 vs x64 builds... 
 
     # If there was no explicit argument saying what to do, default
     # to update, build and test (for native builds).
@@ -2691,6 +2688,8 @@ def main():
     #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
@@ -2707,8 +2706,9 @@ def main():
     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)