feat: add Windows ARM64 (MSVC) build support

.github/workflows/05-windows-build.yml

@@ -18,7 +18,16 @@ jobs:
       matrix:
         include:
           - platform: windows-2022
+            msvc_arch: x64
+            python_version: '3.10'
           - platform: windows-2025
+            msvc_arch: x64
+            python_version: '3.10'
+          # Windows ARM64: Python 3.10 has no official ARM64 installer;
+          # 3.11 is the first CPython release with a Windows-on-ARM build.
+          - platform: windows-11-arm
+            msvc_arch: arm64
+            python_version: '3.11'
 
     env:
       SCCACHE_GHA_ENABLED: "true"
@@ -47,14 +56,14 @@ jobs:
       - name: Set up Python
         uses: actions/setup-python@v7
         with:
-          python-version: '3.10'
+          python-version: ${{ matrix.python_version }}
           cache: 'pip'
           cache-dependency-path: 'pyproject.toml'
 
       - name: Set up MSVC environment
         uses: ilammy/msvc-dev-cmd@v1.13.0
         with:
-          arch: x64
+          arch: ${{ matrix.msvc_arch }}
 
       - name: Set up environment variables
         run: |

src/ailego/CMakeLists.txt

@@ -90,29 +90,35 @@ if(NOT ANDROID AND AUTO_DETECT_ARCH)
         )
     endforeach()
     elseif (HOST_ARCH MATCHES "^(arm|arm64)$")
-        if(MSVC)
-            return()
-        endif()
-        set(MATH_MARCH_FLAG_NEON "-march=armv8-a")
-
-        file(GLOB_RECURSE MATH_FILES_NEON
-          ${CMAKE_CURRENT_SOURCE_DIR}/math/*_dispatch.cc
-          ${CMAKE_CURRENT_SOURCE_DIR}/math/*_dispatch.c
-          ${CMAKE_CURRENT_SOURCE_DIR}/math_batch/*_dispatch.cc
-          ${CMAKE_CURRENT_SOURCE_DIR}/math_batch/*_dispatch.c
-          ${CMAKE_CURRENT_SOURCE_DIR}/math/*_neon.cc
-          ${CMAKE_CURRENT_SOURCE_DIR}/math/*_neon.c
-          ${CMAKE_CURRENT_SOURCE_DIR}/math_batch/*_neon.cc
-          ${CMAKE_CURRENT_SOURCE_DIR}/math_batch/*_neon.c
-        )
+        if(NOT MSVC)
+            set(MATH_MARCH_FLAG_NEON "-march=armv8-a")
+
+            file(GLOB_RECURSE MATH_FILES_NEON
+              ${CMAKE_CURRENT_SOURCE_DIR}/math/*_dispatch.cc
+              ${CMAKE_CURRENT_SOURCE_DIR}/math/*_dispatch.c
+              ${CMAKE_CURRENT_SOURCE_DIR}/math_batch/*_dispatch.cc
+              ${CMAKE_CURRENT_SOURCE_DIR}/math_batch/*_dispatch.c
+              ${CMAKE_CURRENT_SOURCE_DIR}/math/*_neon.cc
+              ${CMAKE_CURRENT_SOURCE_DIR}/math/*_neon.c
+              ${CMAKE_CURRENT_SOURCE_DIR}/math_batch/*_neon.cc
+              ${CMAKE_CURRENT_SOURCE_DIR}/math_batch/*_neon.c
+            )
 
-        foreach(MATH_FILE ${MATH_FILES_NEON})
-          set_source_files_properties(
-              ${MATH_FILE}
-              PROPERTIES
-              COMPILE_FLAGS "${MATH_MARCH_FLAG_NEON}"
-          )
-        endforeach()
+            foreach(MATH_FILE ${MATH_FILES_NEON})
+              set_source_files_properties(
+                  ${MATH_FILE}
+                  PROPERTIES
+                  COMPILE_FLAGS "${MATH_MARCH_FLAG_NEON}"
+              )
+            endforeach()
+        else()
+            # MSVC on ARM64: NEON is the ARMv8 baseline and is always enabled,
+            # so no `-march` flag is required (MSVC does not accept GCC-style
+            # `-march=` anyway). The NEON math kernels still get compiled via
+            # the ALL_SRCS glob above; their guards use the AILEGO_HAVE_NEON /
+            # AILEGO_ARM64 feature macros (defined in platform.h), which resolve
+            # to true on MSVC ARM64 so the bodies actually emit code.
+        endif()
     endif()
 endif()
 

src/ailego/internal/cpu_features.cc

@@ -14,12 +14,11 @@
 
 #include "cpu_features.h"
 #include <cstddef>
+#include <zvec/ailego/internal/platform.h>
 
-#if defined(_MSC_VER)
+#if defined(_MSC_VER) && (defined(_M_IX86) || defined(_M_X64))
 #include <intrin.h>
-#endif
-
-#if (defined(__x86_64__) || defined(__i386__)) && !defined(_MSC_VER)
+#elif defined(__x86_64__) || defined(__i386__)
 #include <cpuid.h>
 #endif
 
@@ -43,7 +42,7 @@ namespace internal {
 
 CpuFeatures::CpuFlags CpuFeatures::flags_;
 
-#if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86))
+#if defined(_MSC_VER) && (defined(_M_IX86) || defined(_M_X64))
 CpuFeatures::CpuFlags::CpuFlags(void)
     : L1_ECX(0), L1_EDX(0), L7_EBX(0), L7_ECX(0), L7_EDX(0) {
   int l1[4] = {0, 0, 0, 0};
@@ -347,7 +346,11 @@ bool CpuFeatures::HYPERVISOR(void) {
 bool CpuFeatures::NEON(void) {
 #if defined(__aarch64__) && defined(__linux__)
   return !!(getauxval(AT_HWCAP) & HWCAP_ASIMD);
-#elif defined(__ARM_NEON)
+#elif defined(AILEGO_HAVE_NEON)
+  // NEON is part of the ARMv8 baseline, so it is unconditionally present when
+  // the target has it (including MSVC ARM64, which predefines only _M_ARM64).
+  // This must agree with Intrinsics()/version.i, which report "Neon" from the
+  // same macro.
   return true;
 #else
   return false;
@@ -356,7 +359,7 @@ bool CpuFeatures::NEON(void) {
 
 const char *CpuFeatures::Intrinsics(void) {
   return ""
-#if defined(__ARM_NEON)
+#if defined(AILEGO_HAVE_NEON)
          "Neon"
 #if defined(__ARM_FEATURE_CRC32)
          "+CRC"

src/ailego/math/distance_matrix_accum_fp32.i

@@ -30,7 +30,7 @@
                                             _mm512_castps_si512(b)))
 #endif  // __AVX512DQ__
 
-#if defined(__ARM_NEON) && !defined(__aarch64__)
+#if defined(AILEGO_HAVE_NEON) && !defined(AILEGO_ARM64)
 static inline float32_t vaddvq_f32(float32x4_t v) {
   float32x2_t s = vadd_f32(vget_low_f32(v), vget_high_f32(v));
   return vget_lane_f32(vpadd_f32(s, s), 0);
@@ -40,13 +40,13 @@ static inline int32_t vaddvq_s32(int32x4_t v) {
   int32x2_t s = vadd_s32(vget_low_s32(v), vget_high_s32(v));
   return vget_lane_s32(vpadd_s32(s, s), 0);
 }
-#endif  //__ARM_NEON && !__aarch64__
+#endif  // AILEGO_HAVE_NEON && !AILEGO_ARM64
 
-#if defined(__aarch64__)
+#if defined(AILEGO_ARM64)
 #define ACCUM_FP32_2X1_NEON ACCUM_FP32_2X1_NEON_A64
 #else
 #define ACCUM_FP32_2X1_NEON ACCUM_FP32_2X1_NEON_A32
-#endif  // __aarch64__
+#endif  // AILEGO_ARM64
 
 //! Compute the distance between matrix and query (FP32, M=2, N=1)
 #define ACCUM_FP32_2X1_SSE(m, q, dim, out, _NORM)                       \

src/ailego/math/distance_matrix_fp32.i

@@ -26,9 +26,9 @@
   _mm256_insertf128_ps(_mm256_castps128_ps256(b), (a), 1)
 #endif  // __AVX__
 
-#if defined(__ARM_NEON) && !defined(__aarch64__)
+#if defined(AILEGO_HAVE_NEON) && !defined(AILEGO_ARM64)
 #define vdupq_laneq_f32(a, b) vdupq_n_f32(vgetq_lane_f32(a, b))
-#endif  // __ARM_NEON && __aarch64__
+#endif  // AILEGO_HAVE_NEON && !AILEGO_ARM64
 
 //! Iterative process of computing distance (FP32, M=2, N=1)
 #define MATRIX_FP32_ITER_2X1_SSE(m, q, _RES, _LOAD, _PROC)         \

src/ailego/math/euclidean_distance_matrix_fp32_dispatch.cc

@@ -18,7 +18,7 @@
 namespace zvec {
 namespace ailego {
 
-#if defined(__ARM_NEON)
+#if defined(AILEGO_HAVE_NEON)
 void SquaredEuclideanDistanceFp32NEON(const float *lhs, const float *rhs,
                                       size_t size, float *out);
 #endif
@@ -49,7 +49,7 @@ void SquaredEuclideanDistanceMatrix<float, 1, 1>::Compute(const ValueType *m,
                                                           const ValueType *q,
                                                           size_t dim,
                                                           float *out) {
-#if defined(__ARM_NEON)
+#if defined(AILEGO_HAVE_NEON)
   SquaredEuclideanDistanceFp32NEON(m, q, dim, out);
 #else
 #if defined(__AVX512F__)
@@ -72,7 +72,7 @@ void SquaredEuclideanDistanceMatrix<float, 1, 1>::Compute(const ValueType *m,
   }
 #endif  // __SSE__
   *out = SquaredEuclideanDistanceFp32Scalar(m, q, dim);
-#endif  // __ARM_NEON
+#endif  // AILEGO_HAVE_NEON
 }
 
 //-----------------------------------------------------------

src/ailego/math/euclidean_distance_matrix_fp32_neon.cc

@@ -19,7 +19,7 @@
 namespace zvec {
 namespace ailego {
 
-#if defined(__ARM_NEON)
+#if defined(AILEGO_HAVE_NEON)
 //! Squared Euclidean Distance
 void SquaredEuclideanDistanceFp32NEON(const float *lhs, const float *rhs,
                                       size_t size, float *out) {
@@ -56,7 +56,7 @@ void SquaredEuclideanDistanceFp32NEON(const float *lhs, const float *rhs,
   *out = result;
 }
 
-#endif  // __ARM_NEON
+#endif  // AILEGO_HAVE_NEON
 
 }  // namespace ailego
 }  // namespace zvec

src/ailego/math/fht_avx2.cc

@@ -12,6 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <zvec/ailego/internal/platform.h>
+
 #if defined(__AVX2__)
 
 #include <immintrin.h>

src/ailego/math/fht_avx512.cc

@@ -12,6 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <zvec/ailego/internal/platform.h>
+
 #if defined(__AVX512F__)
 
 #include <immintrin.h>

src/ailego/math/fht_dispatch.cc

@@ -13,6 +13,7 @@
 // limitations under the License.
 
 #include <ailego/internal/cpu_features.h>
+#include <zvec/ailego/internal/platform.h>
 #include "fht.h"
 
 namespace zvec {
@@ -41,7 +42,7 @@ void fht_kacs_walk_avx512(float *data, size_t len);
 void fht_inv_kacs_walk_avx512(float *data, size_t len);
 void fht_inplace_avx512(float *data, size_t n);
 #endif
-#if defined(__ARM_NEON) && defined(__aarch64__)
+#if defined(AILEGO_ARM64_NEON)
 void fht_flip_sign_neon(const uint8_t *flip, float *data, size_t dim);
 void fht_kacs_walk_neon(float *data, size_t len);
 void fht_inv_kacs_walk_neon(float *data, size_t len);
@@ -52,7 +53,7 @@ void fht_inv_kacs_walk_neon(float *data, size_t len);
 // ============================================================================
 
 void fht_flip_sign(const uint8_t *flip, float *data, size_t dim) {
-#if defined(__ARM_NEON) && defined(__aarch64__)
+#if defined(AILEGO_ARM64_NEON)
   fht_flip_sign_neon(flip, data, dim);
 #else
 #if defined(__AVX512F__)
@@ -75,11 +76,11 @@ void fht_flip_sign(const uint8_t *flip, float *data, size_t dim) {
   }
 #endif
   fht_flip_sign_scalar(flip, data, dim);
-#endif  // __ARM_NEON
+#endif  // AILEGO_ARM64_NEON
 }
 
 void fht_kacs_walk(float *data, size_t len) {
-#if defined(__ARM_NEON) && defined(__aarch64__)
+#if defined(AILEGO_ARM64_NEON)
   fht_kacs_walk_neon(data, len);
 #else
 #if defined(__AVX512F__)
@@ -101,11 +102,11 @@ void fht_kacs_walk(float *data, size_t len) {
   }
 #endif
   fht_kacs_walk_scalar(data, len);
-#endif  // __ARM_NEON
+#endif  // AILEGO_ARM64_NEON
 }
 
 void fht_inv_kacs_walk(float *data, size_t len) {
-#if defined(__ARM_NEON) && defined(__aarch64__)
+#if defined(AILEGO_ARM64_NEON)
   fht_inv_kacs_walk_neon(data, len);
 #else
 #if defined(__AVX512F__)
@@ -127,7 +128,7 @@ void fht_inv_kacs_walk(float *data, size_t len) {
   }
 #endif
   fht_inv_kacs_walk_scalar(data, len);
-#endif  // __ARM_NEON
+#endif  // AILEGO_ARM64_NEON
 }
 
 void fht_inplace(float *data, size_t n) {

src/ailego/math/fht_neon.cc

@@ -12,7 +12,9 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#if defined(__ARM_NEON) && defined(__aarch64__)
+#include <zvec/ailego/internal/platform.h>
+
+#if defined(AILEGO_ARM64_NEON)
 
 #include <cmath>
 #include <cstddef>
@@ -40,7 +42,10 @@ void fht_flip_sign_neon(const uint8_t *flip, float *data, size_t dim) {
     uint32_t b1 = (bits16 >> 1) & 1u;
     uint32_t b2 = (bits16 >> 2) & 1u;
     uint32_t b3 = (bits16 >> 3) & 1u;
-    uint32x4_t bit_mask = {b0, b1, b2, b3};
+    // Build the lane mask via vld1q_u32: MSVC's uint32x4_t is a union type,
+    // so GCC/Clang-style brace initialization of a vector is not portable.
+    const uint32_t bits[4] = {b0, b1, b2, b3};
+    uint32x4_t bit_mask = vld1q_u32(bits);
     uint32x4_t sign_mask = vmulq_u32(bit_mask, sign_bit);
     float32x4_t v = vld1q_f32(&data[i]);
     v = vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(v), sign_mask));
@@ -104,4 +109,4 @@ void fht_inv_kacs_walk_neon(float *data, size_t len) {
 }  // namespace ailego
 }  // namespace zvec
 
-#endif  // __ARM_NEON && __aarch64__
+#endif  // AILEGO_ARM64_NEON

src/ailego/math/fht_sse.cc

@@ -12,6 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <zvec/ailego/internal/platform.h>
+
 #if defined(__SSE2__)
 
 #include <emmintrin.h>

src/ailego/math/inner_product_matrix_fp32_dispatch.cc

@@ -20,7 +20,7 @@ namespace ailego {
 //--------------------------------------------------
 // Dense
 //--------------------------------------------------
-#if defined(__ARM_NEON)
+#if defined(AILEGO_HAVE_NEON)
 float InnerProductFp32NEON(const float *lhs, const float *rhs, size_t size);
 float MinusInnerProductFp32NEON(const float *lhs, const float *rhs,
                                 size_t size);
@@ -49,7 +49,7 @@ float MinusInnerProductFp32Scalar(const float *lhs, const float *rhs,
 //! Compute the distance between matrix and query (FP32, M=1, N=1)
 void InnerProductMatrix<float, 1, 1>::Compute(const float *m, const float *q,
                                               size_t dim, float *out) {
-#if defined(__ARM_NEON)
+#if defined(AILEGO_HAVE_NEON)
   *out = InnerProductFp32NEON(m, q, dim);
 #else
 #if defined(__AVX512F__)
@@ -73,14 +73,14 @@ void InnerProductMatrix<float, 1, 1>::Compute(const float *m, const float *q,
   }
 #endif  // __SSE__
   *out = InnerProductFp32Scalar(m, q, dim);
-#endif  // __ARM_NEON
+#endif  // AILEGO_HAVE_NEON
 }
 
 //! Compute the distance between matrix and query (FP32, M=1, N=1)
 void MinusInnerProductMatrix<float, 1, 1>::Compute(const float *m,
                                                    const float *q, size_t dim,
                                                    float *out) {
-#if defined(__ARM_NEON)
+#if defined(AILEGO_HAVE_NEON)
   *out = MinusInnerProductFp32NEON(m, q, dim);
 #else
 #if defined(__AVX512F__)
@@ -104,7 +104,7 @@ void MinusInnerProductMatrix<float, 1, 1>::Compute(const float *m,
   }
 #endif  // __SSE__
   *out = MinusInnerProductFp32Scalar(m, q, dim);
-#endif  // __ARM_NEON
+#endif  // AILEGO_HAVE_NEON
 }
 
 //--------------------------------------------------