Preserve CUDA internal-test builds without the Windows CMake cycle

.github/workflows/android.yml

@@ -79,8 +79,8 @@ jobs:
       run: |
         set -e -x
         BINARY_SIZE_THRESHOLD_ARGS=""
-        echo "Binary size threshold in bytes: 1597440"
-        BINARY_SIZE_THRESHOLD_ARGS="--threshold_size_in_bytes 1597440"
+        echo "Binary size threshold in bytes: 1599488"
+        BINARY_SIZE_THRESHOLD_ARGS="--threshold_size_in_bytes 1599488"
 
         # Ensure ANDROID_NDK_HOME is available and get its real path
         if [ -z "$ANDROID_NDK_HOME" ]; then

.github/workflows/windows_arm64_python.yml

@@ -11,6 +11,9 @@ on:
       - rel-*
   workflow_dispatch:
 
+permissions:
+  contents: read
+
 concurrency:
   group: ${{ github.workflow }}-${{ github.event_name == 'pull_request' && github.ref || github.sha }}
   cancel-in-progress: true

.github/workflows/windows_cuda.yml

@@ -108,15 +108,53 @@ jobs:
           $buildDir = Join-Path ${{ runner.temp }} "build"
           echo "OnnxRuntimeBuildDirectory=$buildDir" >> $env:GITHUB_ENV
 
-      - name: Build and Clean Binaries
+      - name: Configure build
         working-directory: ${{ runner.temp }}
         run: |
           npm install -g typescript
           if ($lastExitCode -ne 0) {
             exit $lastExitCode
           }
-          # Execute the build process
-          python.exe ${{ github.workspace }}\tools\ci_build\build.py --update --build --config RelWithDebInfo --build_dir build --skip_submodule_sync --build_csharp --parallel --nvcc_threads 4 --flash_nvcc_threads 4 --use_binskim_compliant_compile_flags --cmake_generator "Visual Studio 17 2022" --build_shared_lib --build_wheel --build_java --use_cuda --cuda_home="$env:RUNNER_TEMP\v12.8" --enable_cuda_profiling --use_vcpkg --use_vcpkg_ms_internal_asset_cache --enable_transformers_tool_test --cmake_extra_defines onnxruntime_QUICK_BUILD=ON --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=86 --cmake_extra_defines onnxruntime_ENABLE_CUDA_EP_INTERNAL_TESTS=ON
+          python.exe ${{ github.workspace }}\tools\ci_build\build.py --update --config RelWithDebInfo --build_dir build --skip_submodule_sync --build_csharp --parallel --nvcc_threads 4 --flash_nvcc_threads 4 --use_binskim_compliant_compile_flags --cmake_generator "Visual Studio 17 2022" --build_shared_lib --build_wheel --build_java --use_cuda --cuda_home="$env:RUNNER_TEMP\v12.8" --enable_cuda_profiling --use_vcpkg --use_vcpkg_ms_internal_asset_cache --enable_transformers_tool_test --cmake_extra_defines onnxruntime_QUICK_BUILD=ON --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=86 --cmake_extra_defines onnxruntime_ENABLE_CUDA_EP_INTERNAL_TESTS=ON
+          if ($lastExitCode -ne 0) {
+            exit $lastExitCode
+          }
+        shell: pwsh
+
+      - name: Verify targeted provider-test build
+        shell: pwsh
+        run: |
+          python.exe ${{ github.workspace }}\tools\ci_build\test_gen_test_exports.py
+          if ($LASTEXITCODE -ne 0) {
+            exit $LASTEXITCODE
+          }
+
+          $buildDir = Join-Path $env:OnnxRuntimeBuildDirectory "RelWithDebInfo"
+          $binaryDir = Join-Path $buildDir "RelWithDebInfo"
+          $testBinaries = @("onnxruntime_provider_test.exe", "onnxruntime_provider_test.lib", "onnxruntime_providers_cuda_ut.dll")
+          foreach ($binary in $testBinaries) {
+            $path = Join-Path $binaryDir $binary
+            if (Test-Path -LiteralPath $path) {
+              throw "The targeted-build check requires fresh outputs, but $path already exists."
+            }
+          }
+
+          cmake --build "$buildDir" --config RelWithDebInfo --target onnxruntime_provider_test --parallel
+          if ($LASTEXITCODE -ne 0) {
+            exit $LASTEXITCODE
+          }
+
+          foreach ($binary in $testBinaries) {
+            $path = Join-Path $binaryDir $binary
+            if (-not (Test-Path -LiteralPath $path -PathType Leaf)) {
+              throw "Building only onnxruntime_provider_test did not produce $path"
+            }
+          }
+
+      - name: Build and Clean Binaries
+        working-directory: ${{ runner.temp }}
+        run: |
+          python.exe ${{ github.workspace }}\tools\ci_build\build.py --build --config RelWithDebInfo --build_dir build --skip_submodule_sync --build_csharp --parallel --nvcc_threads 4 --flash_nvcc_threads 4 --use_binskim_compliant_compile_flags --cmake_generator "Visual Studio 17 2022" --build_shared_lib --build_wheel --build_java --use_cuda --cuda_home="$env:RUNNER_TEMP\v12.8" --enable_cuda_profiling --use_vcpkg --use_vcpkg_ms_internal_asset_cache --enable_transformers_tool_test --cmake_extra_defines onnxruntime_QUICK_BUILD=ON --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=86 --cmake_extra_defines onnxruntime_ENABLE_CUDA_EP_INTERNAL_TESTS=ON
           if ($lastExitCode -ne 0) {
             exit $lastExitCode
           }
@@ -230,6 +268,24 @@ jobs:
         shell: pwsh
         run: nvidia-smi
 
+      - name: Verify CUDA EP internal tests execute
+        working-directory: ${{ runner.temp }}\build\RelWithDebInfo\RelWithDebInfo
+        shell: pwsh
+        run: |
+          if (Test-Path -LiteralPath cuda_ep_internal_ci.xml) {
+            Remove-Item -LiteralPath cuda_ep_internal_ci.xml
+          }
+          .\onnxruntime_provider_test.exe --gtest_filter=CUDA_EP_Unittest.All --gtest_output=xml:cuda_ep_internal_ci.xml
+          if ($LASTEXITCODE -ne 0) {
+            exit $LASTEXITCODE
+          }
+
+          [xml]$results = Get-Content -LiteralPath cuda_ep_internal_ci.xml -Raw
+          $executedTest = $results.SelectNodes("/testsuites/testsuite[@name='CUDA_EP_Unittest']/testcase[@name='All' and @status='run' and @result='completed']")
+          if ($executedTest.Count -ne 1) {
+            throw "CUDA_EP_Unittest.All did not execute: the test must be present and must not be skipped."
+          }
+
       - name: Run Tests
         working-directory: ${{ runner.temp }}
         run: |

cmake/CMakeLists.txt

@@ -70,11 +70,6 @@ option(onnxruntime_ENABLE_PYTHON "Enable python bindings" OFF)
 option(onnxruntime_ENABLE_MEMLEAK_CHECKER "Experimental: Enable memory leak checker in Windows debug build" OFF)
 option(onnxruntime_ENABLE_CONVSYMKERNELAVX2_SAT_CHECKER "Experimental: Enable ConvSymKernelAvx2 assembly saturation checker in build" OFF)
 option(onnxruntime_USE_CUDA "Build with CUDA support" OFF)
-# Enable ONNX Runtime CUDA EP's internal unit tests that directly access the EP's internal functions instead of through
-# OpKernels. When the option is ON, we will have two copies of GTest library in the same process. It is not a typical
-# use. If you hit any problem with that, please do not report it to GTest. Turn OFF the following build option instead.
-cmake_dependent_option(onnxruntime_ENABLE_CUDA_EP_INTERNAL_TESTS "Build with CUDA unit tests" OFF "onnxruntime_USE_CUDA;onnxruntime_BUILD_UNIT_TESTS" OFF)
-
 cmake_dependent_option(onnxruntime_USE_CUDA_NHWC_OPS "Build CUDA with NHWC op support" ON "onnxruntime_USE_CUDA" OFF)
 cmake_dependent_option(onnxruntime_BUILD_CUDA_EP_AS_PLUGIN "Build CUDA EP as a separate plugin shared library instead of the legacy in-tree provider" OFF "onnxruntime_USE_CUDA" OFF)
 option(onnxruntime_BUILD_CUDA_QUANT_PREPROCESS "Build CUDA weight-packing module onnxruntime_cuda_quant_preprocess.so" OFF)
@@ -97,6 +92,12 @@ option(onnxruntime_USE_ARM_NEON_NCHWC "Build with ARM Neon NCHWc kernels in MLAS
 option(onnxruntime_USE_KLEIDIAI "Build with KleidiAI integration in MLAS" OFF)
 option(onnxruntime_USE_QMX_KLEIDIAI_COEXIST "Build with QMX and Arm KLEIDIAI libraries" OFF)
 option(onnxruntime_BUILD_UNIT_TESTS "Build ONNXRuntime unit tests" ON)
+# Declare both prerequisites before evaluating this dependent option on a fresh configure.
+# Enable ONNX Runtime CUDA EP's internal unit tests that directly access the EP's internal functions instead of through
+# OpKernels. When the option is ON, we will have two copies of GTest library in the same process. It is not a typical
+# use. If you hit any problem with that, please do not report it to GTest. Turn OFF the following build option instead.
+cmake_dependent_option(onnxruntime_ENABLE_CUDA_EP_INTERNAL_TESTS "Build with CUDA unit tests" OFF "onnxruntime_USE_CUDA;onnxruntime_BUILD_UNIT_TESTS" OFF)
+
 # Materialize the ONNX node-test corpus from ONNX's Python generators into the build tree at
 # configure/build time, instead of depending on the on-disk corpus shipped in the ONNX source
 # archive. This detaches ORT from ONNX PR #7959 (which deletes onnx/backend/test/data/node).

cmake/onnxruntime_test_exports.cmake

@@ -0,0 +1,57 @@
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT License.
+
+function(onnxruntime_export_test_symbols host)
+  cmake_parse_arguments(EXPORTS "" "OBJECT_TARGET" "HOST_LIBS;MODULE_LIBS" ${ARGN})
+  find_package(Python COMPONENTS Interpreter REQUIRED)
+
+  set(module_files $<TARGET_OBJECTS:${EXPORTS_OBJECT_TARGET}>)
+  set(host_files)
+  set(library_dependencies)
+  foreach(side HOST MODULE)
+    foreach(library IN LISTS EXPORTS_${side}_LIBS)
+      if(TARGET ${library})
+        get_target_property(library_type ${library} TYPE)
+        if(library_type STREQUAL "STATIC_LIBRARY" OR library_type STREQUAL "UNKNOWN_LIBRARY")
+          set(library_files $<TARGET_FILE:${library}>)
+        elseif(library_type STREQUAL "OBJECT_LIBRARY")
+          set(library_files $<TARGET_OBJECTS:${library}>)
+        else()
+          continue()
+        endif()
+        if(side STREQUAL "HOST")
+          list(APPEND host_files ${library_files})
+        else()
+          list(APPEND module_files ${library_files})
+        endif()
+        list(APPEND library_dependencies ${library})
+      endif()
+    endforeach()
+  endforeach()
+  if(NOT host_files)
+    message(FATAL_ERROR "No static host libraries found for ${host}'s test exports")
+  endif()
+  list(REMOVE_DUPLICATES host_files)
+  list(REMOVE_DUPLICATES module_files)
+  list(REMOVE_DUPLICATES library_dependencies)
+
+  set(export_dir "${CMAKE_CURRENT_BINARY_DIR}/${host}_exports/$<CONFIG>")
+  file(GENERATE OUTPUT "${export_dir}/host.rsp" CONTENT "\"$<JOIN:${host_files},\"\n\">\"\n")
+  file(GENERATE OUTPUT "${export_dir}/module.rsp" CONTENT "\"$<JOIN:${module_files},\"\n\">\"\n")
+  set(export_script "${REPO_ROOT}/tools/ci_build/gen_test_exports.py")
+  set(export_file "${export_dir}/exports.def")
+  set(force_include_file "${export_dir}/force_include.cc")
+  add_custom_command(OUTPUT "${export_file}" "${force_include_file}"
+    COMMAND ${Python_EXECUTABLE} "${export_script}"
+      --linker "${CMAKE_LINKER}"
+      --host "${export_dir}/host.rsp"
+      --module "${export_dir}/module.rsp"
+      --output "${export_file}"
+      --force-include "${force_include_file}"
+    DEPENDS "${export_script}" "${export_dir}/host.rsp" "${export_dir}/module.rsp"
+      ${EXPORTS_OBJECT_TARGET} ${library_dependencies} ${host_files} ${module_files}
+    VERBATIM)
+  set_target_properties(${host} PROPERTIES ENABLE_EXPORTS ON WINDOWS_EXPORT_ALL_SYMBOLS OFF)
+  set_source_files_properties("${force_include_file}" PROPERTIES SKIP_PRECOMPILE_HEADERS ON)
+  target_sources(${host} PRIVATE "${export_file}" "${force_include_file}")
+endfunction()

cmake/onnxruntime_test_pch.cmake

@@ -6,7 +6,7 @@ if(CMAKE_CXX_COMPILER_ID MATCHES "MSVC")
     "${CMAKE_CURRENT_SOURCE_DIR}/test_pch.h"
   )
   if (TARGET onnxruntime_provider_test)
-    target_precompile_headers(onnxruntime_provider_test PRIVATE
+    target_precompile_headers(${onnxruntime_provider_test_target} PRIVATE
       "${CMAKE_CURRENT_SOURCE_DIR}/test_pch.h"
     )
   endif()