Update WebGPU

.clang-tidy

@@ -163,7 +163,7 @@ Checks: >
 # keep-sorted end
 WarningsAsErrors: '*'
 HeaderFilterRegex: '.*'
-ExcludeHeaderFilterRegex: '\.fbs\.h|common/cmdline\.h|/mini_\w*\.h'
+ExcludeHeaderFilterRegex: '\.fbs\.h|common/cmdline\.h|mini_[^/]*\.h'
 FormatStyle: 'file'
 CheckOptions:
     -   key: modernize-use-default-member-init.UseAssignment

.github/workflows/testing-macos.yml

@@ -105,3 +105,42 @@ jobs:
           -destination 'generic/platform=iOS Simulator'
           build
         working-directory: apps/HelloiOS
+
+  webgpu:
+    if: "!contains(github.event.pull_request.labels.*.name, 'skip_buildbots')"
+    name: arm-64 / host-webgpu
+    runs-on: macos-26
+
+    steps:
+      - uses: actions/checkout@v7
+
+      - uses: astral-sh/setup-uv@v7
+
+      - name: Sync CI environment
+        run: |
+          uv sync --python 'cpython-3.10.20-macos-aarch64-none' --group ci-llvm-main --no-install-project
+          echo "${GITHUB_WORKSPACE}/.venv/bin" >> "$GITHUB_PATH"
+          echo "VIRTUAL_ENV=${GITHUB_WORKSPACE}/.venv" >> "$GITHUB_ENV"
+
+      - name: Configure LLVM paths
+        run: echo "Halide_LLVM_ROOT=$(halide-llvm --prefix)" >> "$GITHUB_ENV"
+
+      - name: Cache vcpkg binary archives
+        uses: actions/cache@v4
+        with:
+          path: ~/.cache/vcpkg/archives
+          key: vcpkg-archives-macos-arm64-webgpu-${{ hashFiles('vcpkg.json', 'vcpkg-configuration.json', 'cmake/vcpkg-triplets/arm64-osx-halide.cmake', 'cmake/vcpkg-ports/**') }}
+
+      - name: Configure CMake
+        run: cmake --preset ci-macos-arm-64-webgpu -DHalide_TARGET=host-webgpu
+
+      - name: Build
+        run: cmake --build build
+
+      - name: Test (host-webgpu)
+        run: ctest --test-dir build --output-on-failure -j "$(sysctl -n hw.ncpu)" -R correctness_gpu
+        env:
+          # CMake's find_package resolves Dawn fine at configure/build time, but the
+          # JIT's runtime dlopen (JITModule.cpp's load_webgpu) is a separate lookup
+          # that needs to be told explicitly where the vcpkg-installed dylib lives.
+          HL_WEBGPU_NATIVE_LIB: ${{ github.workspace }}/build/vcpkg_installed/arm64-osx-halide/lib/libwebgpu_dawn.dylib

CMakePresets.json

@@ -57,6 +57,15 @@
       "name": "ci-macos-arm-64",
       "inherits": "ci"
     },
+    {
+      "name": "ci-macos-arm-64-webgpu",
+      "inherits": "ci-macos-arm-64",
+      "cacheVariables": {
+        "VCPKG_TARGET_TRIPLET": "arm64-osx-halide",
+        "VCPKG_MANIFEST_FEATURES": "developer;webgpu-native",
+        "VCPKG_CHAINLOAD_TOOLCHAIN_FILE": "${sourceDir}/cmake/toolchain.ci-macos-arm-64.cmake"
+      }
+    },
     {
       "name": "ci-macos-x86-64",
       "inherits": "ci"

Makefile

@@ -911,7 +911,8 @@ RUNTIME_CPP_COMPONENTS = \
   vulkan \
   wasm_cpu_features \
   wasm_thread_id \
-  webgpu_dawn \
+  webgpu_dawn_arm \
+  webgpu_dawn_x86 \
   webgpu_emscripten \
   windows_aarch64_cpu_features_arm \
   windows_clock \
@@ -1111,6 +1112,8 @@ RUNTIME_TRIPLE_WIN_GENERIC_64 = "x86_64-unknown-windows-unknown"
 
 RUNTIME_TRIPLE_WEBGPU_32 = "wasm32-unknown-unknown-unknown"
 RUNTIME_TRIPLE_WEBGPU_64 = "wasm64-unknown-unknown-unknown"
+RUNTIME_TRIPLE_WEBGPU_ARM_64 = "aarch64-unknown-unknown-unknown"
+RUNTIME_TRIPLE_WEBGPU_X86_64 = "x86_64-unknown-unknown-unknown"
 
 # `-fno-threadsafe-statics` is very important here (note that it allows us to use a 'modern' C++
 # standard but still skip threadsafe guards for static initialization in our runtime code)
@@ -1164,6 +1167,22 @@ $(BUILD_DIR)/initmod.windows_%_64.ll: $(SRC_DIR)/runtime/windows_%.cpp
 	@mkdir -p $(@D)
 	$(CLANG) $(CXX_WARNING_FLAGS) $(RUNTIME_CXX_FLAGS) -m64 -target $(RUNTIME_TRIPLE_WIN_GENERIC_64) -fshort-wchar -DCOMPILING_HALIDE_RUNTIME -DBITS_64 -emit-llvm -S $(SRC_DIR)/runtime/windows_$*.cpp -o $@ -MMD -MP -MF $(BUILD_DIR)/initmod.windows_$*_64.d
 
+$(BUILD_DIR)/initmod.webgpu_dawn_arm_32.ll: $(SRC_DIR)/runtime/webgpu_dawn_arm.cpp
+	@mkdir -p $(@D)
+	$(CLANG) $(CXX_WARNING_FLAGS) $(RUNTIME_CXX_FLAGS) -m32 -target $(RUNTIME_TRIPLE_32) -DCOMPILING_HALIDE_RUNTIME -DBITS_32 -emit-llvm -S $(SRC_DIR)/runtime/webgpu_dawn_arm.cpp -o $@ -MMD -MP -MF $(BUILD_DIR)/initmod.webgpu_dawn_arm_32.d
+
+$(BUILD_DIR)/initmod.webgpu_dawn_arm_64.ll: $(SRC_DIR)/runtime/webgpu_dawn_arm.cpp
+	@mkdir -p $(@D)
+	$(CLANG) $(CXX_WARNING_FLAGS) $(RUNTIME_CXX_FLAGS) -m64 -target $(RUNTIME_TRIPLE_WEBGPU_ARM_64) -DCOMPILING_HALIDE_RUNTIME -DBITS_64 -emit-llvm -S $(SRC_DIR)/runtime/webgpu_dawn_arm.cpp -o $@ -MMD -MP -MF $(BUILD_DIR)/initmod.webgpu_dawn_arm_64.d
+
+$(BUILD_DIR)/initmod.webgpu_dawn_x86_32.ll: $(SRC_DIR)/runtime/webgpu_dawn_x86.cpp
+	@mkdir -p $(@D)
+	$(CLANG) $(CXX_WARNING_FLAGS) $(RUNTIME_CXX_FLAGS) -m32 -target $(RUNTIME_TRIPLE_32) -DCOMPILING_HALIDE_RUNTIME -DBITS_32 -emit-llvm -S $(SRC_DIR)/runtime/webgpu_dawn_x86.cpp -o $@ -MMD -MP -MF $(BUILD_DIR)/initmod.webgpu_dawn_x86_32.d
+
+$(BUILD_DIR)/initmod.webgpu_dawn_x86_64.ll: $(SRC_DIR)/runtime/webgpu_dawn_x86.cpp
+	@mkdir -p $(@D)
+	$(CLANG) $(CXX_WARNING_FLAGS) $(RUNTIME_CXX_FLAGS) -m64 -target $(RUNTIME_TRIPLE_WEBGPU_X86_64) -DCOMPILING_HALIDE_RUNTIME -DBITS_64 -emit-llvm -S $(SRC_DIR)/runtime/webgpu_dawn_x86.cpp -o $@ -MMD -MP -MF $(BUILD_DIR)/initmod.webgpu_dawn_x86_64.d
+
 $(BUILD_DIR)/initmod.webgpu_%_32.ll: $(SRC_DIR)/runtime/webgpu_%.cpp
 	@mkdir -p $(@D)
 	$(CLANG) $(CXX_WARNING_FLAGS) $(RUNTIME_CXX_FLAGS) -m32 -target $(RUNTIME_TRIPLE_WEBGPU_32) -DCOMPILING_HALIDE_RUNTIME -DBITS_32 -emit-llvm -S $(SRC_DIR)/runtime/webgpu_$*.cpp -o $@ -MMD -MP -MF $(BUILD_DIR)/initmod.webgpu_$*_32.d
@@ -1172,6 +1191,22 @@ $(BUILD_DIR)/initmod.webgpu_%_64.ll: $(SRC_DIR)/runtime/webgpu_%.cpp
 	@mkdir -p $(@D)
 	$(CLANG) $(CXX_WARNING_FLAGS) $(RUNTIME_CXX_FLAGS) -m64 -target $(RUNTIME_TRIPLE_WEBGPU_64) -DCOMPILING_HALIDE_RUNTIME -DBITS_64 -emit-llvm -S $(SRC_DIR)/runtime/webgpu_$*.cpp -o $@ -MMD -MP -MF $(BUILD_DIR)/initmod.webgpu_$*_64.d
 
+$(BUILD_DIR)/initmod.webgpu_dawn_arm_32_debug.ll: $(SRC_DIR)/runtime/webgpu_dawn_arm.cpp
+	@mkdir -p $(@D)
+	$(CLANG) $(CXX_WARNING_FLAGS) -g -DDEBUG_RUNTIME $(RUNTIME_CXX_FLAGS) -m32 -target $(RUNTIME_TRIPLE_32) -DCOMPILING_HALIDE_RUNTIME -DBITS_32 -emit-llvm -S $(SRC_DIR)/runtime/webgpu_dawn_arm.cpp -o $@ -MMD -MP -MF $(BUILD_DIR)/initmod.webgpu_dawn_arm_32_debug.d
+
+$(BUILD_DIR)/initmod.webgpu_dawn_arm_64_debug.ll: $(SRC_DIR)/runtime/webgpu_dawn_arm.cpp
+	@mkdir -p $(@D)
+	$(CLANG) $(CXX_WARNING_FLAGS) -g -DDEBUG_RUNTIME $(RUNTIME_CXX_FLAGS) -m64 -target $(RUNTIME_TRIPLE_WEBGPU_ARM_64) -DCOMPILING_HALIDE_RUNTIME -DBITS_64 -emit-llvm -S $(SRC_DIR)/runtime/webgpu_dawn_arm.cpp -o $@ -MMD -MP -MF $(BUILD_DIR)/initmod.webgpu_dawn_arm_64_debug.d
+
+$(BUILD_DIR)/initmod.webgpu_dawn_x86_32_debug.ll: $(SRC_DIR)/runtime/webgpu_dawn_x86.cpp
+	@mkdir -p $(@D)
+	$(CLANG) $(CXX_WARNING_FLAGS) -g -DDEBUG_RUNTIME $(RUNTIME_CXX_FLAGS) -m32 -target $(RUNTIME_TRIPLE_32) -DCOMPILING_HALIDE_RUNTIME -DBITS_32 -emit-llvm -S $(SRC_DIR)/runtime/webgpu_dawn_x86.cpp -o $@ -MMD -MP -MF $(BUILD_DIR)/initmod.webgpu_dawn_x86_32_debug.d
+
+$(BUILD_DIR)/initmod.webgpu_dawn_x86_64_debug.ll: $(SRC_DIR)/runtime/webgpu_dawn_x86.cpp
+	@mkdir -p $(@D)
+	$(CLANG) $(CXX_WARNING_FLAGS) -g -DDEBUG_RUNTIME $(RUNTIME_CXX_FLAGS) -m64 -target $(RUNTIME_TRIPLE_WEBGPU_X86_64) -DCOMPILING_HALIDE_RUNTIME -DBITS_64 -emit-llvm -S $(SRC_DIR)/runtime/webgpu_dawn_x86.cpp -o $@ -MMD -MP -MF $(BUILD_DIR)/initmod.webgpu_dawn_x86_64_debug.d
+
 $(BUILD_DIR)/initmod.webgpu_%_32_debug.ll: $(SRC_DIR)/runtime/webgpu_%.cpp
 	@mkdir -p $(@D)
 	$(CLANG) $(CXX_WARNING_FLAGS) -g -DDEBUG_RUNTIME $(RUNTIME_CXX_FLAGS) -m32 -target $(RUNTIME_TRIPLE_WEBGPU_32) -DCOMPILING_HALIDE_RUNTIME -DBITS_32 -emit-llvm -S $(SRC_DIR)/runtime/webgpu_$*.cpp -o $@ -MMD -MP -MF $(BUILD_DIR)/initmod.webgpu_$*_32_debug.d

cmake/FindHalide_WebGPU.cmake

@@ -7,16 +7,77 @@ if (EXISTS "$ENV{HL_WEBGPU_NATIVE_LIB}")
     set(Halide_WebGPU_NATIVE_LIB "$ENV{HL_WEBGPU_NATIVE_LIB}" CACHE FILEPATH "")
 endif ()
 
-find_library(Halide_WebGPU_NATIVE_LIB NAMES webgpu_dawn wgpu)
+# Try to find Dawn via vcpkg's CMake config first
+if (NOT TARGET Halide::WebGPU)
+    # Look for DawnConfig.cmake in vcpkg installed directory
+    if (DEFINED ENV{VCPKG_ROOT})
+        set(_dawn_cmake_path "$ENV{VCPKG_ROOT}/installed/${VCPKG_TARGET_TRIPLET}/share/dawn")
+        if (EXISTS "${_dawn_cmake_path}/DawnConfig.cmake")
+            # Use find_package to properly resolve dependencies
+            find_package(Dawn CONFIG QUIET)
+            if (TARGET dawn::webgpu_dawn AND NOT TARGET Halide::WebGPU)
+                add_library(Halide::WebGPU ALIAS dawn::webgpu_dawn)
+            elseif (TARGET dawn::dawn_public_config AND NOT TARGET Halide::WebGPU)
+                get_target_property(_dawn_loc dawn::dawn_public_config IMPORTED_LOCATION)
+                if (_dawn_loc)
+                    add_library(Halide::WebGPU UNKNOWN IMPORTED)
+                    set_target_properties(Halide::WebGPU
+                        PROPERTIES
+                        IMPORTED_LOCATION "${_dawn_loc}"
+                    )
+                    get_target_property(
+                        _dawn_link_deps dawn::dawn_public_config INTERFACE_LINK_LIBRARIES
+                    )
+                    if (_dawn_link_deps)
+                        set_target_properties(Halide::WebGPU
+                            PROPERTIES
+                            INTERFACE_LINK_LIBRARIES "${_dawn_link_deps}"
+                        )
+                    endif ()
+                endif ()
+            endif ()
+        elseif (EXISTS "${_dawn_cmake_path}/DawnTargets.cmake")
+            include("${_dawn_cmake_path}/DawnTargets.cmake")
+            if (TARGET dawn::webgpu_dawn AND NOT TARGET Halide::WebGPU)
+                add_library(Halide::WebGPU ALIAS dawn::webgpu_dawn)
+            endif ()
+        endif ()
+    endif ()
+endif ()
+
+# If Dawn wasn't found via vcpkg config, fall back to manual find
+if (NOT TARGET Halide::WebGPU)
+    # Prefer shared library for runtime loading (dlopen)
+    find_library(
+        Halide_WebGPU_NATIVE_LIB
+        NAMES webgpu_dawn wgpu
+        PATHS "${CMAKE_CURRENT_LIST_DIR}/../vcpkg_installed/${VCPKG_TARGET_TRIPLET}/lib"
+        NO_DEFAULT_PATH
+    )
+    if (NOT Halide_WebGPU_NATIVE_LIB)
+        find_library(Halide_WebGPU_NATIVE_LIB NAMES webgpu_dawn wgpu)
+    endif ()
+
+    include(FindPackageHandleStandardArgs)
+    find_package_handle_standard_args(
+        Halide_WebGPU
+        REQUIRED_VARS Halide_WebGPU_NATIVE_LIB
+        HANDLE_COMPONENTS
+    )
 
-include(FindPackageHandleStandardArgs)
-find_package_handle_standard_args(
-    Halide_WebGPU
-    REQUIRED_VARS Halide_WebGPU_NATIVE_LIB
-    HANDLE_COMPONENTS
-)
+    if (Halide_WebGPU_NATIVE_LIB)
+        add_library(Halide::WebGPU UNKNOWN IMPORTED)
+        set_target_properties(Halide::WebGPU
+            PROPERTIES
+            IMPORTED_LOCATION "${Halide_WebGPU_NATIVE_LIB}"
+        )
 
-if (Halide_WebGPU_NATIVE_LIB AND NOT TARGET Halide::WebGPU)
-    add_library(Halide::WebGPU UNKNOWN IMPORTED)
-    set_target_properties(Halide::WebGPU PROPERTIES IMPORTED_LOCATION "${Halide_WebGPU_NATIVE_LIB}")
+        if (APPLE)
+            set_target_properties(Halide::WebGPU
+                PROPERTIES
+                INTERFACE_LINK_LIBRARIES
+                    "-framework Cocoa;-framework IOKit;-framework Foundation;-framework IOSurface;-framework QuartzCore;-framework Metal"
+            )
+        endif ()
+    endif ()
 endif ()

cmake/vcpkg-triplets/arm64-osx-halide.cmake

@@ -0,0 +1,16 @@
+set(VCPKG_TARGET_ARCHITECTURE arm64)
+set(VCPKG_CRT_LINKAGE dynamic)
+set(VCPKG_LIBRARY_LINKAGE static)
+set(VCPKG_CMAKE_SYSTEM_NAME Darwin)
+set(VCPKG_OSX_ARCHITECTURES arm64)
+
+# Dawn must be a shared library so it can be loaded via dlopen at runtime.
+if (PORT STREQUAL "dawn")
+    set(VCPKG_LIBRARY_LINKAGE dynamic)
+endif ()
+
+# Skip Debug builds in CI: they're unused (we only ever load Release Dawn) and
+# ~2x the build time / ~30x the artifact size for no benefit.
+if (DEFINED ENV{CI})
+    set(VCPKG_BUILD_TYPE release)
+endif ()

cmake/vcpkg-triplets/x64-osx-halide.cmake

@@ -0,0 +1,15 @@
+set(VCPKG_TARGET_ARCHITECTURE x64)
+set(VCPKG_CRT_LINKAGE dynamic)
+set(VCPKG_LIBRARY_LINKAGE static)
+set(VCPKG_CMAKE_SYSTEM_NAME Darwin)
+
+# Dawn must be a shared library so it can be loaded via dlopen at runtime.
+if (PORT STREQUAL "dawn")
+    set(VCPKG_LIBRARY_LINKAGE dynamic)
+endif ()
+
+# Skip Debug builds in CI: they're unused (we only ever load Release Dawn) and
+# ~2x the build time / ~30x the artifact size for no benefit.
+if (DEFINED ENV{CI})
+    set(VCPKG_BUILD_TYPE release)
+endif ()