Merge branch 'main' into fork/vawale/add-cmake-for-onnx-app

alexreinking · alexreinking · commit b3240d2e0d52 · 2025-08-05T08:08:22.000-04:00
diff --git a/.github/workflows/pip.yml b/.github/workflows/pip.yml
@@ -59,13 +59,6 @@ jobs:
           fetch-depth: 0
           fetch-tags: true
 
-      # See: https://github.com/pypa/setuptools-scm/issues/455
-      - name: Suppress git version tag
-        if: github.event_name == 'push' && github.ref_name == 'main'
-        run: |
-          echo 'local_scheme = "no-local-version"' >> pyproject.toml
-          git update-index --assume-unchanged pyproject.toml
-
       - uses: ilammy/msvc-dev-cmd@v1
       - uses: lukka/get-cmake@v3.28.4
 
@@ -210,12 +203,17 @@ jobs:
         env:
           CIBW_BUILD: "cp3*-${{ matrix.platform_tag }}"
           CIBW_SKIP: "cp3{5,6,7,8}*"
+          # Suppress the git version tag (necessary for TestPyPI)
+          CIBW_ENVIRONMENT: >
+            SETUPTOOLS_SCM_OVERRIDES_FOR_HALIDE='{local_scheme="no-local-version"}'
           CIBW_ENVIRONMENT_MACOS: >
             CMAKE_PREFIX_PATH='${{ github.workspace }}/opt'
             Python_ROOT_DIR=''
+            SETUPTOOLS_SCM_OVERRIDES_FOR_HALIDE='{local_scheme="no-local-version"}'
           CIBW_ENVIRONMENT_WINDOWS: >
             CMAKE_GENERATOR=Ninja
             CMAKE_PREFIX_PATH='${{ github.workspace }}\opt'
+            SETUPTOOLS_SCM_OVERRIDES_FOR_HALIDE='{local_scheme="no-local-version"}'
           CIBW_MANYLINUX_X86_64_IMAGE: "ghcr.io/halide/manylinux_2_28_x86_64-llvm:${{ env.LLVM_VERSION }}"
           CIBW_TEST_COMMAND: >
             cmake -G Ninja -S {project}/python_bindings/apps -B build -DCMAKE_BUILD_TYPE=Release &&
diff --git a/.github/workflows/presubmit.yml b/.github/workflows/presubmit.yml
@@ -8,6 +8,8 @@ on:
       - '**.h'
       - '**.c'
       - '**.cpp'
+      - '**CMakeLists.txt'
+      - '**.cmake'
       - 'run-clang-tidy.sh'
       - 'run-clang-format.sh'
       - '.github/workflows/presubmit.yml'
diff --git a/apps/hannk/CMakeLists.txt b/apps/hannk/CMakeLists.txt
@@ -29,8 +29,8 @@ set(CMAKE_CXX_EXTENSIONS NO)
 # (We need to do this even if HANNK_BUILD_TFLITE is off,
 # so that the .tflite file parser can get the right schema)
 set(TFLITE_VERSION_MAJOR "2" CACHE STRING "Major version of TFLite to assume")
-set(TFLITE_VERSION_MINOR "8" CACHE STRING "Minor version of TFLite to assume")
-set(TFLITE_VERSION_PATCH "3" CACHE STRING "Patch version of TFLite to assume")
+set(TFLITE_VERSION_MINOR "19" CACHE STRING "Minor version of TFLite to assume")
+set(TFLITE_VERSION_PATCH "0" CACHE STRING "Patch version of TFLite to assume")
 set(TFLITE_VERSION "${TFLITE_VERSION_MAJOR}.${TFLITE_VERSION_MINOR}.${TFLITE_VERSION_PATCH}")
 
 # ----------------------------
diff --git a/apps/hannk/Makefile b/apps/hannk/Makefile
@@ -59,8 +59,8 @@ clean:
 # ---------------------- TFLite glue
 
 TFLITE_VERSION_MAJOR ?= 2
-TFLITE_VERSION_MINOR ?= 8
-TFLITE_VERSION_PATCH ?= 3
+TFLITE_VERSION_MINOR ?= 19
+TFLITE_VERSION_PATCH ?= 0
 
 TFLITE_VERSION = $(TFLITE_VERSION_MAJOR).$(TFLITE_VERSION_MINOR).$(TFLITE_VERSION_PATCH)
 TFLITE_TAG = v$(TFLITE_VERSION)
diff --git a/apps/hannk/tflite/CMakeLists.txt b/apps/hannk/tflite/CMakeLists.txt
@@ -20,49 +20,49 @@ set(FLATBUFFERS_BUILD_FLATC OFF)
 # Enable this to see details about downloading -- useful for debugging
 # set(FETCHCONTENT_QUIET NO)
 
-FetchContent_Declare(tflite
-                     GIT_REPOSITORY https://github.com/tensorflow/tensorflow
-                     GIT_TAG ${TFLITE_TAG}
-                     GIT_SHALLOW TRUE)
+# tflite includes neon2sse transitively, but it's pinned to an old version
+# that declares a minimum CMake version of 3.0. We override it here to a
+# new enough version that CMake 4+ won't refuse to build it.
+FetchContent_Declare(
+    neon2sse
+    GIT_REPOSITORY https://github.com/intel/ARM_NEON_2_x86_SSE
+    GIT_TAG 4732aac1e6c02984a12635d85c4644c2ffe585ca
+    GIT_SHALLOW TRUE
+)
+
+FetchContent_Declare(
+    tflite
+    GIT_REPOSITORY https://github.com/tensorflow/tensorflow
+    GIT_TAG ${TFLITE_TAG}
+    GIT_SHALLOW TRUE
+    SOURCE_SUBDIR tensorflow/lite/c
+)
+
+block(SCOPE_FOR POLICIES VARIABLES)
+    # Suppress warnings about rotting CMake code we don't control
+    set(CMAKE_POLICY_DEFAULT_CMP0135 OLD) # DOWNLOAD_EXTRACT_TIMESTAMP
+    set(CMAKE_POLICY_DEFAULT_CMP0169 OLD) # FetchContent_Populate
+    set(CMAKE_POLICY_DEFAULT_CMP0177 OLD) # install() path normalization
+
+    # Configuration for tflite + upstream deps
+    set(ABSL_PROPAGATE_CXX_STD ON)
+
+    # Suppress warnings in tflite code
+    add_compile_options(
+        $<$<CXX_COMPILER_ID:Clang,AppleClang>:-Wno-anon-enum-enum-conversion>
+        $<$<CXX_COMPILER_ID:Clang,AppleClang>:-Wno-deprecated-this-capture>
+        $<$<CXX_COMPILER_ID:Clang,AppleClang>:-Wno-shadow-uncaptured-local>
+        $<$<CXX_COMPILER_ID:Clang,AppleClang>:-Wno-shadow>
+        $<$<CXX_COMPILER_ID:Clang,AppleClang>:-Wno-tautological-type-limit-compare>
+        $<$<CXX_COMPILER_ID:Clang,AppleClang>:-Wno-unused-template>
+    )
+
+    FetchContent_MakeAvailable(tflite)
+
+    set_property(TARGET tensorflowlite_c PROPERTY EXCLUDE_FROM_ALL TRUE)
+endblock()
 
 FetchContent_GetProperties(tflite)
-if (NOT tflite_POPULATED)
-    FetchContent_Populate(tflite)
-    # Some of the subprojects (e.g. Eigen) are very noisy and emit status messages all the time.
-    # Temporary ignore status messages while adding this to silence it. Ugly but effective.
-    set(OLD_CMAKE_MESSAGE_LOG_LEVEL ${CMAKE_MESSAGE_LOG_LEVEL})
-    set(CMAKE_MESSAGE_LOG_LEVEL WARNING)
-    add_subdirectory(${tflite_SOURCE_DIR}/tensorflow/lite/c ${tflite_BINARY_DIR})
-    set(CMAKE_MESSAGE_LOG_LEVEL ${OLD_CMAKE_MESSAGE_LOG_LEVEL})
-endif ()
-
-# tensorflowlite_c is implicitly declared by this FetchContent.
-# Mark it as EXCLUDE_FROM_ALL so that it won't be built unless we actually
-# depend on it (which we might not depending on HANNK_BUILD_TFLITE)
-set_property(TARGET tensorflowlite_c PROPERTY EXCLUDE_FROM_ALL TRUE)
-
-# Disable some noisy warnings in abseil
-foreach (LIB IN ITEMS
-            absl_base
-            absl_graphcycles_internal
-            absl_malloc_internal
-            absl_synchronization
-            absl_time
-            absl_time_zone)
-    target_compile_options(${LIB}
-                           PRIVATE
-                           $<$<CXX_COMPILER_ID:Clang,AppleClang>:-Wno-tautological-type-limit-compare>
-                           $<$<CXX_COMPILER_ID:Clang,AppleClang>:-Wno-unused-template>
-                           $<$<CXX_COMPILER_ID:Clang,AppleClang>:-Wno-shadow>
-                           $<$<CXX_COMPILER_ID:Clang,AppleClang>:-Wno-shadow-uncaptured-local>
-                           $<$<CXX_COMPILER_ID:Clang,AppleClang>:-Wno-anon-enum-enum-conversion>)
-endforeach ()
-
-# Disable some noisy warnings in tflite
-target_compile_options(tensorflow-lite
-                       PRIVATE
-                       $<$<CXX_COMPILER_ID:Clang,AppleClang>:-Wno-gnu-inline-cpp-without-extern>
-                       $<$<CXX_COMPILER_ID:GNU>:-Wno-ignored-attributes>)
 
 # Make an interface library that is just to get the tflite headers,
 # without any implied linkage
diff --git a/cmake/HalideGeneratorHelpers.cmake b/cmake/HalideGeneratorHelpers.cmake
@@ -68,10 +68,6 @@ function(add_halide_generator TARGET)
             add_custom_target("${ARG_PACKAGE_NAME}")
         endif ()
 
-        if (NOT Halide_FOUND)
-            find_package(Halide REQUIRED)
-        endif ()
-
         if (ARG_SOURCES MATCHES ".py$")
             if (ARG_LINK_LIBRARIES)
                 message(FATAL_ERROR "You cannot specify LINK_LIBRARIES in conjunction with Python source code.")
@@ -94,6 +90,10 @@ function(add_halide_generator TARGET)
         else ()
             add_executable(${TARGET} ${ARG_SOURCES})
             add_executable(${gen} ALIAS ${TARGET})
+
+            if (NOT TARGET Halide::Generator)
+                find_package(Halide REQUIRED)
+            endif ()
             target_link_libraries(${TARGET} PRIVATE Halide::Generator ${ARG_LINK_LIBRARIES})
 
             _Halide_place_dll(${TARGET})
diff --git a/packaging/pip/TrampolineConfig.cmake.in b/packaging/pip/TrampolineConfig.cmake.in
@@ -1,6 +1,9 @@
 cmake_minimum_required(VERSION 3.28)
 
 include(CMakeFindDependencyMacro)
-find_dependency(Python 3)
+find_dependency(
+    Python 3
+    COMPONENTS Interpreter Development.Module
+)
 
 include("${Python_SITEARCH}/halide/@INSTALL_DIR@/@PACKAGE@Config.cmake")
diff --git a/pyproject.toml b/pyproject.toml
@@ -2,6 +2,7 @@
 requires = [
     "pybind11>=2.11.1",
     "scikit-build-core==0.11.5",
+    "setuptools-scm>=8.3.1",
 ]
 build-backend = "scikit_build_core.build"
 
@@ -66,6 +67,9 @@ dev = [
     "scikit-build-core==0.11.5",
     "setuptools-scm>=8.3.1",
 ]
+apps = [
+    "onnx"
+]
 tools = [
     "cmake>=3.28",
     "ninja>=1.11",
@@ -111,6 +115,9 @@ if.platform-system = "^win32"
 inherit.cmake.define = "append"
 cmake.define.Halide_WASM_BACKEND = "OFF"
 
+[tool.setuptools_scm]
+# Needs to exist for scikit-build-core to use setuptools_scm
+
 [tool.ruff.lint]
 # docs: https://docs.astral.sh/ruff/rules/
 select = [
@@ -165,7 +172,3 @@ search = "#define HALIDE_VERSION_MINOR {current_version}"
 src = "src/runtime/HalideRuntime.h"
 version_template = "{patch}"
 search = "#define HALIDE_VERSION_PATCH {current_version}"
-
-# Must be last table in file since pip.yml appends to it
-# See: https://github.com/pypa/setuptools-scm/issues/455
-[tool.setuptools_scm]
diff --git a/requirements.txt b/requirements.txt
@@ -1,5 +1,5 @@
 # This file was autogenerated by uv via the following command:
-#    uv export --all-groups --no-hashes --no-emit-project --format requirements-txt
+#    uv export --all-groups --no-hashes --no-emit-project -o requirements.txt
 cli-ui==0.19.0
     # via tbump
 cmake==4.0.3
@@ -18,14 +18,18 @@ numpy==2.0.2 ; python_full_version < '3.10'
     # via
     #   halide
     #   imageio
+    #   onnx
 numpy==2.2.6 ; python_full_version == '3.10.*'
     # via
     #   halide
     #   imageio
-numpy==2.3.1 ; python_full_version >= '3.11'
+    #   onnx
+numpy==2.3.2 ; python_full_version >= '3.11'
     # via
     #   halide
     #   imageio
+    #   onnx
+onnx==1.18.0
 packaging==25.0
     # via
     #   scikit-build-core
@@ -34,8 +38,10 @@ pathspec==0.12.1
     # via scikit-build-core
 pillow==11.3.0
     # via imageio
+protobuf==6.31.1
+    # via onnx
 pybind11==3.0.0
-ruff==0.12.4
+ruff==0.12.7
 schema==0.7.7
     # via tbump
 scikit-build-core==0.11.5
@@ -51,9 +57,10 @@ tomli==2.2.1 ; python_full_version < '3.11'
     #   setuptools-scm
 tomlkit==0.11.8
     # via tbump
-typing-extensions==4.14.1 ; python_full_version < '3.11'
+typing-extensions==4.14.1
     # via
     #   exceptiongroup
+    #   onnx
     #   setuptools-scm
 unidecode==1.4.0
     # via cli-ui
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
@@ -551,7 +551,7 @@ if (WITH_SERIALIZATION)
     add_custom_command(
         OUTPUT "${fb_header}"
         COMMAND flatbuffers::flatc --cpp --cpp-std C++17 --no-union-value-namespacing --keep-prefix --filename-suffix ".fbs" -o "${fb_dir}" "${fb_def}"
-        DEPENDS "${fb_def}"
+        DEPENDS flatbuffers::flatc "${fb_def}"
         VERBATIM
     )
     add_custom_target(generate_fb_header DEPENDS "${fb_header}")
diff --git a/src/LLVM_Runtime_Linker.cpp b/src/LLVM_Runtime_Linker.cpp
@@ -890,6 +890,13 @@ std::unique_ptr<llvm::Module> link_with_wasm_jit_runtime(llvm::LLVMContext *c, c
     constexpr bool allow_stripping_all_weak_functions = true;
     link_modules(modules, t, allow_stripping_all_weak_functions);
 
+    // The initmods are compiled by clang and have some attributes associated with
+    // them that are irrelevant to Wasm. Strip them out.
+    for (llvm::Function &F : *modules[0]) {
+        F.removeFnAttr("target-features");
+        F.removeFnAttr("target-cpu");
+    }
+
     return std::move(modules[0]);
 }
 
diff --git a/src/runtime/CMakeLists.txt b/src/runtime/CMakeLists.txt
@@ -280,14 +280,14 @@ foreach (i IN LISTS RUNTIME_CPP)
             else()
                 add_custom_command(OUTPUT "${LL}"
                                    COMMAND ${CMAKE_CROSSCOMPILING_EMULATOR} ${CMAKE_C_COMPILER_LAUNCHER} $<TARGET_FILE:clang> ${clang_flags} -o "${LL}" "$<SHELL_PATH:${SOURCE}>"
-                                   DEPENDS "${SOURCE}"
+                                   DEPENDS $<TARGET_NAME:clang> "${SOURCE}"
                                    DEPFILE "${basename}.d"
                                    VERBATIM)
             endif()
 
             add_custom_command(OUTPUT "${BC}"
                                COMMAND ${CMAKE_CROSSCOMPILING_EMULATOR} $<TARGET_FILE:llvm-as> "${LL}" -o "${BC}"
-                               DEPENDS "${LL}"
+                               DEPENDS $<TARGET_NAME:llvm-as> "${LL}"
                                VERBATIM)
 
             add_custom_command(OUTPUT "${INITMOD}"
@@ -314,7 +314,7 @@ foreach (i IN LISTS RUNTIME_LL)
 
     add_custom_command(OUTPUT "${BC}"
                        COMMAND ${CMAKE_CROSSCOMPILING_EMULATOR} $<TARGET_FILE:llvm-as> "${LL_TRANSFORMED}" -o "${BC}"
-                       DEPENDS "${LL_TRANSFORMED}"
+                       DEPENDS $<TARGET_NAME:llvm-as> "${LL_TRANSFORMED}"
                        VERBATIM)
     add_custom_command(OUTPUT "${INITMOD}"
                        COMMAND binary2cpp "halide_internal_initmod_${i}_ll" < "${BC}" > "${INITMOD}"
diff --git a/test/correctness/simd_op_check_wasm.cpp b/test/correctness/simd_op_check_wasm.cpp
@@ -80,7 +80,7 @@ class SimdOpCheckWASM : public SimdOpCheckTest {
 
                 // Create vector with identical lanes
                 // (Note that later LLVMs will use 64-bit constants for some smaller splats)
-                check("i8x16.splat", 16 * w, u8_1 * u8(42));
+                check("i8x16.splat", 16 * w, u8_1 * u8(in_u8(2) << 1));
                 // LLVM13 likes to emit all of these as v128.const
                 check("v128.const", 8 * w, u16_1 * u16(42));
                 check("v128.const", 4 * w, u32_1 * u32(42));
diff --git a/test/generator/CMakeLists.txt b/test/generator/CMakeLists.txt
@@ -80,7 +80,7 @@ function(add_wasm_executable TARGET)
 
     add_custom_command(OUTPUT "${TARGET}.wasm" "${TARGET}.js"
                        COMMAND ${EMCC} ${EMCC_FLAGS} ${INCLUDES} ${SRCS} ${DEPS} -o "${TARGET}.js"
-                       DEPENDS ${SRCS} ${DEPS}
+                       DEPENDS ${EMCC} ${SRCS} ${DEPS}
                        VERBATIM)
 
     add_custom_target("${TARGET}" ALL

Original file line number	Diff line number	Diff line change
`@@ -551,7 +551,7 @@ if (WITH_SERIALIZATION)`
`551`	`551`	`add_custom_command(`
`552`	`552`	`OUTPUT "${fb_header}"`
`553`	`553`	`COMMAND flatbuffers::flatc --cpp --cpp-std C++17 --no-union-value-namespacing --keep-prefix --filename-suffix ".fbs" -o "${fb_dir}" "${fb_def}"`
`554`		`- DEPENDS "${fb_def}"`
	`554`	`+ DEPENDS flatbuffers::flatc "${fb_def}"`
`555`	`555`	`VERBATIM`
`556`	`556`	`)`
`557`	`557`	`add_custom_target(generate_fb_header DEPENDS "${fb_header}")`