Merge branch 'master' into sync_msft_05082025

Author: Jaswanth51

.github/actions/macos-ci-setup/action.yml

@@ -3,9 +3,8 @@ description: "Common setup steps for macOS CI pipelines"
 
 inputs:
   platform_machine:
-    required: false
+    required: true
     type: string
-    default: "arm64"
   python_version:
     required: false
     type: string

.github/workflows/mac.yml

@@ -23,32 +23,42 @@ jobs:
     uses: ./.github/workflows/macos-ci-build-and-test-workflow.yml
     with:
       # Only build arm64 for CPU
-      matrix_exclude: >-
+      matrix_include: >-
         [
-          {"platform_machine": "x86_64"}
+          {"machine": "arm64", "target": "arm64", "build_config": "Debug"},
+          {"machine": "arm64", "target": "arm64", "build_config": "Release"}
         ]
 
   coreml:
     uses: ./.github/workflows/macos-ci-build-and-test-workflow.yml
     with:
       use_coreml: true
-
+      matrix_include: >-
+        [
+          {"machine": "x86_64", "target": "x86_64", "build_config": "Release"},
+          {"machine": "arm64", "target": "arm64", "build_config": "Debug"},
+          {"machine": "arm64", "target": "arm64", "build_config": "Release"}
+        ]
   xnnpack:
     uses: ./.github/workflows/macos-ci-build-and-test-workflow.yml
     with:
       use_xnnpack: true
       # only build arm64/Debug for XNNPack
-      matrix_exclude: >-
+      matrix_include: >-
         [
-          {"platform_machine": "x86_64"},
-          {"platform_machine": "arm64", "build_config": "Release"}
+          {"machine": "arm64", "target": "arm64", "build_config": "Debug"}
         ]
 
   webgpu:
     uses: ./.github/workflows/macos-ci-build-and-test-workflow.yml
     with:
       use_webgpu: true
-
+      matrix_include: >-
+        [
+          {"machine": "arm64", "target": "x86_64", "build_config": "Release"},
+          {"machine": "arm64", "target": "arm64", "build_config": "Debug"},
+          {"machine": "arm64", "target": "arm64", "build_config": "Release"}
+        ]
   iphone_simulator:
     runs-on: macos-15
 

.github/workflows/macos-ci-build-and-test-workflow.yml

@@ -20,26 +20,37 @@ on:
         required: false
         type: string
         default: "3.11"
-      matrix_exclude:
+      matrix_include:
         required: false
         type: string
-        description: "JSON string specifying combinations to exclude from the matrix"
+        description: "JSON string specifying combinations to include in the matrix"
         # we do not have enough resources to run all combinations
-        # Exclude x86_64 + Debug combination by default
+        # Include arm64 + Debug combination by default
         default: >-
           [
-            {"platform_machine": "x86_64", "build_config": "Debug"}
+            {"machine": "arm64", "target": "arm64", "build_config": "Debug"}
           ]
 
 jobs:
   build-and-test:
     strategy:
       matrix:
-        platform_machine: ["x86_64", "arm64"]
-        build_config: ["Debug", "Release"]
-        exclude: ${{ fromJSON(inputs.matrix_exclude) }}
+        # To avoid creating too many jobs, we explicitly specify the combinations to run.
+        #
+        # # Specify the machine architecture to run the build on.
+        # machine: ["x86_64", "arm64"]
+        #
+        # # Specify the target architecture for the build.
+        # target: ["x86_64", "arm64"]
+        #
+        # # Specify the build configuration for the build.
+        # build_config: ["Debug", "Release"]
+        #
+        include: ${{ fromJSON(inputs.matrix_include) }}
 
-    runs-on: ${{ matrix.platform_machine == 'x86_64' && 'macos-13' || 'macos-15' }}
+    # "macos-13" is a x86_64 image, and "macos-15" is an arm64 image.
+    # see also: https://github.com/actions/runner-images/blob/main/README.md
+    runs-on: ${{ matrix.machine == 'x86_64' && 'macos-13' || 'macos-15' }}
     env:
       build_flags: >
         --build_dir ./build
@@ -53,11 +64,13 @@ jobs:
         --build_wheel
         ${{ inputs.use_webgpu && '--use_webgpu' || '' }}
         ${{ inputs.use_xnnpack && '--use_xnnpack' || '' }}
-        ${{ inputs.use_coreml && '--use_coreml' || '' }}
         ${{ inputs.use_coreml && '--use_coreml --skip_onnx_tests' || '' }}
         --use_vcpkg --use_vcpkg_ms_internal_asset_cache
         --config ${{ matrix.build_config }}
-      xcode_version: ${{ matrix.platform_machine == 'x86_64' && '14.3.1' || '16' }}
+        --osx_arch ${{ matrix.target }}
+
+      # xCode version needs to match the "runs-on" configuration.
+      xcode_version: ${{ matrix.machine == 'x86_64' && '14.3.1' || '16' }}
 
     steps:
       - name: Checkout code
@@ -66,7 +79,7 @@ jobs:
       - name: macOS CI pipeline prepare steps
         uses: ./.github/actions/macos-ci-setup
         with:
-          platform_machine: ${{ matrix.platform_machine }}
+          platform_machine: ${{ matrix.machine }}
           python_version: ${{ inputs.python_version }}
           xcode_version: ${{ env.xcode_version }}
           use_cache: true
@@ -108,6 +121,8 @@ jobs:
           make install DESTDIR=${{ github.workspace }}/build/installed
 
       - name: Running Tests (build.py --test)
+        # Skip tests when cross-compiling
+        if: ${{ matrix.machine == matrix.target }}
         shell: bash
         working-directory: ${{ github.workspace }}
         run: |

cmake/CMakeLists.txt

@@ -558,6 +558,7 @@ else()
   check_cxx_compiler_flag(-Wambiguous-reversed-operator HAS_AMBIGUOUS_REVERSED_OPERATOR)
   # -Winterference-size was added in GCC 13
   check_cxx_compiler_flag(-Winterference-size HAS_INTERFERENCE_SIZE)
+  check_cxx_compiler_flag(-Warray-bounds HAS_ARRAY_BOUNDS)
   check_cxx_compiler_flag(-Wbitwise-instead-of-logical HAS_BITWISE_INSTEAD_OF_LOGICAL)
   check_cxx_compiler_flag(-Wcast-function-type HAS_CAST_FUNCTION_TYPE)
   check_cxx_compiler_flag(-Wcatch-value HAS_CATCH_VALUE)

cmake/external/onnxruntime_external_deps.cmake

@@ -747,7 +747,11 @@ if (onnxruntime_USE_WEBGPU)
           #
           # - (private) Fulfill the BinSkim requirements
           #   Some build warnings are not allowed to be disabled in project level.
-          ${Patch_EXECUTABLE} --binary --ignore-whitespace -p1 < ${PROJECT_SOURCE_DIR}/patches/dawn/dawn_binskim.patch)
+          ${Patch_EXECUTABLE} --binary --ignore-whitespace -p1 < ${PROJECT_SOURCE_DIR}/patches/dawn/dawn_binskim.patch &&
+
+          # Android devices doesn't seem to allow fp16 in uniforms so the WebGPU EP has to manually handle passing an fp32
+          # in the uniform and converting to fp16 before using.
+          ${Patch_EXECUTABLE} --binary --ignore-whitespace -p1 < ${PROJECT_SOURCE_DIR}/patches/dawn/uniform_and_storage_buffer_16_bit_access.patch)
 
       onnxruntime_fetchcontent_declare(
         dawn

cmake/onnxruntime_config.h.in

@@ -3,6 +3,7 @@
 
 #pragma once
 
+#cmakedefine HAS_ARRAY_BOUNDS
 #cmakedefine HAS_BITWISE_INSTEAD_OF_LOGICAL
 #cmakedefine HAS_CAST_FUNCTION_TYPE
 #cmakedefine HAS_CATCH_VALUE

cmake/onnxruntime_webassembly.cmake

@@ -101,6 +101,9 @@ if (NOT onnxruntime_USE_VCPKG)
   target_compile_options(onnx PRIVATE -Wno-unused-parameter -Wno-unused-variable)
 endif()
 
+# Include the Node.js helper for finding and validating Node.js and NPM
+include(node_helper.cmake)
+
 if (onnxruntime_BUILD_WEBASSEMBLY_STATIC_LIB)
     bundle_static_library(onnxruntime_webassembly
       ${PROTOBUF_LIB}
@@ -148,11 +151,6 @@ if (onnxruntime_BUILD_WEBASSEMBLY_STATIC_LIB)
         GTest::gtest
       )
 
-      find_program(NODE_EXECUTABLE node required)
-      if (NOT NODE_EXECUTABLE)
-        message(FATAL_ERROR "Node is required for a test")
-      endif()
-
       add_test(NAME onnxruntime_webassembly_test
         COMMAND ${NODE_EXECUTABLE} onnxruntime_webassembly_test.js
         WORKING_DIRECTORY $<TARGET_FILE_DIR:onnxruntime_webassembly_test>
@@ -343,6 +341,19 @@ else()
     )
   endif()
 
+  #
+  # Apply post-processing script for the generated JavaScript file
+  #
+  list(APPEND onnxruntime_webassembly_script_deps "${ONNXRUNTIME_ROOT}/wasm/wasm_post_build.js")
+  add_custom_command(
+    TARGET onnxruntime_webassembly
+    POST_BUILD
+    # Backup file at $<TARGET_FILE_NAME:onnxruntime_webassembly>.bak
+    COMMAND ${CMAKE_COMMAND} -E copy_if_different "$<TARGET_FILE_NAME:onnxruntime_webassembly>" "$<TARGET_FILE_NAME:onnxruntime_webassembly>.bak"
+    COMMAND ${CMAKE_COMMAND} -E echo "Performing post-process for $<TARGET_FILE_NAME:onnxruntime_webassembly>"
+    COMMAND ${NODE_EXECUTABLE} "${ONNXRUNTIME_ROOT}/wasm/wasm_post_build.js" "$<TARGET_FILE_NAME:onnxruntime_webassembly>"
+  )
+
   set_target_properties(onnxruntime_webassembly PROPERTIES LINK_DEPENDS "${onnxruntime_webassembly_script_deps}")
 
   set(target_name_list ort)
@@ -373,61 +384,4 @@ else()
   endif()
 
   set_target_properties(onnxruntime_webassembly PROPERTIES OUTPUT_NAME ${target_name} SUFFIX ".mjs")
-
-  if (onnxruntime_ENABLE_WEBASSEMBLY_THREADS)
-    #
-    # The following POST_BUILD script is a workaround for enabling:
-    # - using onnxruntime-web with Multi-threading enabled when import from CDN
-    # - using onnxruntime-web when consumed in some frameworks like Vite
-    #
-    # In the use case mentioned above, the file name of the script may be changed. So we need to replace the line:
-    # `new Worker(new URL("ort-wasm-*.mjs", import.meta.url),`
-    # with
-    # `new Worker(new URL(import.meta.url),`
-    #
-    # This behavior is introduced in https://github.com/emscripten-core/emscripten/pull/22165. Since it's unlikely to be
-    # reverted, and there is no config to disable this behavior, we have to use a post-build script to workaround it.
-    #
-
-    # Generate a script to do the post-build work
-    file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/wasm_post_build.js "
-      const fs = require('fs');
-      const path = require('path');
-
-      // node wasm_post_build.js <mjsFilePath>
-      const mjsFilePath = process.argv[2];
-      let contents = fs.readFileSync(mjsFilePath).toString();
-
-      const regex = 'new Worker\\\\(new URL\\\\(\".+?\", ?import\\\\.meta\\\\.url\\\\),';
-      const matches = [...contents.matchAll(new RegExp(regex, 'g'))];
-      if (matches.length !== 1) {
-        throw new Error(
-          `Unexpected number of matches for \"\${regex}\" in \"\${mjsFilePath}\": \${matches.length}.`,
-        );
-      }
-
-      // Replace the only occurrence.
-      contents = contents.replace(
-        new RegExp(regex),
-        `new Worker(new URL(import.meta.url),`,
-      );
-
-      fs.writeFileSync(mjsFilePath, contents);
-    "
-    )
-
-    find_program(NODE_EXECUTABLE node required)
-    if (NOT NODE_EXECUTABLE)
-      message(FATAL_ERROR "Node is required to run the post-build script")
-    endif()
-
-    add_custom_command(
-      TARGET onnxruntime_webassembly
-      POST_BUILD
-      # Backup file at $<TARGET_FILE_NAME:onnxruntime_webassembly>.bak
-      COMMAND ${CMAKE_COMMAND} -E copy_if_different "$<TARGET_FILE_NAME:onnxruntime_webassembly>" "$<TARGET_FILE_NAME:onnxruntime_webassembly>.bak"
-      COMMAND ${CMAKE_COMMAND} -E echo "Performing post-process for $<TARGET_FILE_NAME:onnxruntime_webassembly>"
-      COMMAND ${NODE_EXECUTABLE} "${CMAKE_CURRENT_BINARY_DIR}/wasm_post_build.js" "$<TARGET_FILE_NAME:onnxruntime_webassembly>"
-    )
-  endif()
 endif()

cmake/patches/dawn/uniform_and_storage_buffer_16_bit_access.patch

@@ -0,0 +1,39 @@
+diff --git a/src/dawn/native/vulkan/DeviceVk.cpp b/src/dawn/native/vulkan/DeviceVk.cpp
+index c01d64e40f..0f1f4beae4 100644
+--- a/src/dawn/native/vulkan/DeviceVk.cpp
++++ b/src/dawn/native/vulkan/DeviceVk.cpp
+@@ -464,13 +464,15 @@ ResultOrError<VulkanDeviceKnobs> Device::CreateDevice(VkPhysicalDevice vkPhysica
+         DAWN_ASSERT(usedKnobs.HasExt(DeviceExt::ShaderFloat16Int8) &&
+                     mDeviceInfo.shaderFloat16Int8Features.shaderFloat16 == VK_TRUE &&
+                     usedKnobs.HasExt(DeviceExt::_16BitStorage) &&
+-                    mDeviceInfo._16BitStorageFeatures.storageBuffer16BitAccess == VK_TRUE &&
++                    mDeviceInfo._16BitStorageFeatures.storageBuffer16BitAccess == VK_TRUE /*&&
+                     mDeviceInfo._16BitStorageFeatures.uniformAndStorageBuffer16BitAccess ==
+-                        VK_TRUE);
++                        VK_TRUE*/);
+
+         usedKnobs.shaderFloat16Int8Features.shaderFloat16 = VK_TRUE;
+         usedKnobs._16BitStorageFeatures.storageBuffer16BitAccess = VK_TRUE;
+-        usedKnobs._16BitStorageFeatures.uniformAndStorageBuffer16BitAccess = VK_TRUE;
++        if (mDeviceInfo._16BitStorageFeatures.uniformAndStorageBuffer16BitAccess == VK_TRUE) {
++            usedKnobs._16BitStorageFeatures.uniformAndStorageBuffer16BitAccess = VK_TRUE;
++        }
+         if (mDeviceInfo._16BitStorageFeatures.storageInputOutput16 == VK_TRUE) {
+             usedKnobs._16BitStorageFeatures.storageInputOutput16 = VK_TRUE;
+         }
+diff --git a/src/dawn/native/vulkan/PhysicalDeviceVk.cpp b/src/dawn/native/vulkan/PhysicalDeviceVk.cpp
+index a324c101ed..8d64da750f 100644
+--- a/src/dawn/native/vulkan/PhysicalDeviceVk.cpp
++++ b/src/dawn/native/vulkan/PhysicalDeviceVk.cpp
+@@ -269,8 +269,9 @@ void PhysicalDevice::InitializeSupportedFeaturesImpl() {
+     if (mDeviceInfo.HasExt(DeviceExt::ShaderFloat16Int8) &&
+         mDeviceInfo.HasExt(DeviceExt::_16BitStorage) &&
+         mDeviceInfo.shaderFloat16Int8Features.shaderFloat16 == VK_TRUE &&
+-        mDeviceInfo._16BitStorageFeatures.storageBuffer16BitAccess == VK_TRUE &&
+-        mDeviceInfo._16BitStorageFeatures.uniformAndStorageBuffer16BitAccess == VK_TRUE) {
++        mDeviceInfo._16BitStorageFeatures.storageBuffer16BitAccess == VK_TRUE /*&&
++        WebGPU EP needs to ensure we don't put fp16 values in uniforms when this patch is applied.
++        mDeviceInfo._16BitStorageFeatures.uniformAndStorageBuffer16BitAccess == VK_TRUE*/) {
+         // ONNX Runtime Patch: enable shaderF16 on all devices.
+         EnableFeature(Feature::ShaderF16);
+         shaderF16Enabled = true;