Add Windows ARM64 build and CI support

.github/workflows/ci.yml

@@ -193,14 +193,16 @@ jobs:
           arch: arm64
         - os: macos-15-intel
           arch: x86_64
+        - os: windows-11-arm
+          arch: ARM64
 
     steps:
       - uses: actions/checkout@v4
         with:
           submodules: recursive
 
       - name: Cache MKL and oneDNN (Windows)
-        if: runner.os == 'Windows'
+        if: runner.os == 'Windows' && matrix.arch != 'ARM64'
         uses: actions/cache@v4
         with:
           path: |
@@ -219,7 +221,7 @@ jobs:
           CIBW_ENVIRONMENT_MACOS: "CTRANSLATE2_ROOT='/usr/local' MACOSX_DEPLOYMENT_TARGET=11.00"
           CIBW_BEFORE_ALL_LINUX: python/tools/prepare_build_environment_linux.sh
           CIBW_BEFORE_ALL_MACOS: python/tools/prepare_build_environment_macos.sh
-          CIBW_BEFORE_ALL_WINDOWS: bash python/tools/prepare_build_environment_windows.sh
+          CIBW_BEFORE_ALL_WINDOWS: bash python/tools/${{ matrix.arch == 'ARM64' && 'prepare_build_environment_windows_arm64.sh' || 'prepare_build_environment_windows.sh' }}
           CIBW_BEFORE_BUILD: pip install -r python/install_requirements.txt
           CIBW_MANYLINUX_X86_64_IMAGE: manylinux_2_28
           CIBW_MANYLINUX_AARCH64_IMAGE: manylinux_2_28
@@ -299,6 +301,10 @@ jobs:
             artifact_pattern: python-wheels-macOS-arm64
             wheel_pattern: "*cp312*macosx*arm64.whl"
 
+          - os: windows-11-arm
+            artifact_pattern: python-wheels-Windows-ARM64
+            wheel_pattern: "*cp312*win_arm64.whl"
+
     steps:
       - name: Set up Python 3.12
         uses: actions/setup-python@v5

CMakeLists.txt

@@ -272,7 +272,7 @@ macro(ct2_compile_kernels_for_isa isa flag)
   list(APPEND SOURCES ${CMAKE_CURRENT_BINARY_DIR}/kernels_${isa}.cc)
 endmacro()
 
-if(CMAKE_SYSTEM_PROCESSOR MATCHES "(arm64)|(aarch64)"
+if(CMAKE_SYSTEM_PROCESSOR MATCHES "(arm64)|(aarch64)|(ARM64)"
     OR (APPLE AND CMAKE_OSX_ARCHITECTURES STREQUAL "arm64"))
   add_definitions(-DCT2_ARM64_BUILD)
   set(CT2_BUILD_ARCH "arm64")

python/tests/test_translator.py

@@ -68,7 +68,8 @@ def test_contains_model(tmp_dir):
 def test_get_supported_compute_types():
     compute_types = ctranslate2.get_supported_compute_types("cpu")
     assert "float32" in compute_types
-    assert "int8" in compute_types
+    if "int8" not in compute_types:
+        pytest.skip("int8 compute type is not supported by this CPU backend")
     assert "int8_float32" in compute_types
 
 
@@ -88,6 +89,9 @@ def test_compute_type():
     with pytest.raises(TypeError, match="incompatible constructor arguments"):
         ctranslate2.Translator(model_path, compute_type=["int8", "int16"])
 
+    if "int8" not in ctranslate2.get_supported_compute_types("cpu"):
+        pytest.skip("int8 compute type is not supported by this CPU backend")
+
     translator = ctranslate2.Translator(model_path, compute_type="int8")
     assert translator.compute_type == "int8_float32"
 

python/tools/prepare_build_environment_windows_arm64.sh

@@ -0,0 +1,31 @@
+#! /bin/bash
+
+set -e
+set -x
+set -o pipefail
+
+OPENBLAS_VERSION=0.3.28
+OPENBLAS_ROOT="$PWD/openblas-install"
+
+curl --netrc-optional -L -o openblas.tar.gz https://github.com/OpenMathLib/OpenBLAS/releases/download/v${OPENBLAS_VERSION}/OpenBLAS-${OPENBLAS_VERSION}.tar.gz
+tar xzf openblas.tar.gz
+cd OpenBLAS-${OPENBLAS_VERSION}
+(
+  cmake -G Ninja -DCMAKE_POLICY_VERSION_MINIMUM=3.5 -DTARGET=ARMV8 -DBINARY=64 -DNOFORTRAN=ON -DBUILD_WITHOUT_LAPACK=ON -DONLY_CBLAS=ON -DCMAKE_C_COMPILER=clang-cl -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=ON -DCMAKE_INSTALL_PREFIX="$OPENBLAS_ROOT" -B build .
+  cmake --build build --config Release --target install --parallel
+) > /dev/null 2>&1
+cd ..
+rm -rf OpenBLAS-${OPENBLAS_VERSION} openblas.tar.gz
+
+NPROC=$(nproc)
+
+mkdir build
+cd build
+cmake -DCMAKE_POLICY_VERSION_MINIMUM=3.5 -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=$CTRANSLATE2_ROOT -DWITH_MKL=OFF -DWITH_OPENBLAS=ON -DWITH_RUY=OFF -DOPENMP_RUNTIME=COMP -DOPENBLAS_INCLUDE_DIR="$OPENBLAS_ROOT/include/openblas" -DOPENBLAS_LIBRARY="$OPENBLAS_ROOT/lib/openblas.lib" -DBUILD_CLI=OFF ..
+cmake --build . --config Release --target install --parallel $NPROC --verbose
+cd ..
+rm -r build
+
+cp README.md python/
+cp $CTRANSLATE2_ROOT/bin/ctranslate2.dll python/ctranslate2/
+cp "$OPENBLAS_ROOT/bin/openblas.dll" python/ctranslate2/

src/cpu/vec_neon.h

@@ -160,7 +160,7 @@ namespace ctranslate2 {
       }
 
       static inline value_type round(value_type v) {
-#ifdef __aarch64__
+#if defined(__aarch64__) || defined(_M_ARM64)
         return vrndiq_f32(v);
 #else
         float temp[4] = {std::nearbyintf(v[0]), std::nearbyintf(v[1]), std::nearbyintf(v[2]), std::nearbyintf(v[3])};