Use NEON on Linux arm64

.github/workflows/test.yaml

@@ -24,6 +24,18 @@ jobs:
       # after the upload, so dist/test-unit stays out of the artifact
       - run: make test-unit
       - run: cargo test --locked --manifest-path tests/Cargo.toml
+  test-linux-aarch64-extension:
+    runs-on: ubuntu-24.04-arm
+    steps:
+      - uses: actions/checkout@v4
+      - uses: astral-sh/setup-uv@v3
+        with:
+          enable-cache: true
+      - run: ./scripts/vendor.sh
+      - run: make loadable static
+      - run: uv sync --directory tests
+      - run: make test-loadable python=./tests/.venv/bin/python
+      - run: make test-unit
   build-macos-x86_64-extension:
     runs-on: macos-15-intel
     steps:

CLAUDE.md

@@ -36,7 +36,7 @@ Run `./scripts/vendor.sh` first to download vendored dependencies (sqlite3.c, sh
 - Set `AR=` to use a different archiver (default: ar)
 
 **SIMD control:**
-- SIMD is auto-enabled on Darwin x86_64 (AVX) and Darwin arm64 (NEON)
+- SIMD is auto-enabled on Darwin x86_64 (AVX), Darwin arm64 (NEON), and Linux arm64 (NEON)
 - Set `OMIT_SIMD=1` to disable SIMD optimizations
 
 ### Testing

Makefile

@@ -66,6 +66,16 @@ ifndef OMIT_SIMD
 	ifeq ($(shell uname -sm),Darwin arm64)
 	CFLAGS += -mcpu=apple-m1 -DSQLITE_VEC_ENABLE_NEON
 	endif
+	# NEON is in the arm64 baseline (GCC and Clang define __ARM_NEON for
+	# aarch64 with no -m flag), so the Linux build needs only the define. It
+	# follows the compiler's target, not uname: 32-bit Raspberry Pi OS runs a
+	# 64-bit kernel, so uname says aarch64 while gcc targets 32-bit ARM, which
+	# lacks the AArch64-only intrinsics the NEON kernels use.
+	ifeq ($(shell uname -s),Linux)
+	ifneq ($(filter aarch64-%,$(shell $(CC) -dumpmachine)),)
+	CFLAGS += -DSQLITE_VEC_ENABLE_NEON
+	endif
+	endif
 	endif
 endif
 

sqlite-vec.c

@@ -320,26 +320,32 @@ static i32 l1_int8_neon(const void *pVect1v, const void *pVect2v,
   int32x4_t acc3 = vdupq_n_s32(0);
   int32x4_t acc4 = vdupq_n_s32(0);
 
+  // vabdq_s8 returns |a - b| (up to 255) in signed lanes, so each difference
+  // is reinterpreted as unsigned before the widening pairwise adds
   while (pVect1 < pEnd1 - 63) {
     int8x16_t v1 = vld1q_s8(pVect1);
     int8x16_t v2 = vld1q_s8(pVect2);
-    int8x16_t diff1 = vabdq_s8(v1, v2);
-    acc1 = vaddq_s32(acc1, vpaddlq_u16(vpaddlq_u8(diff1)));
+    uint8x16_t diff1 = vreinterpretq_u8_s8(vabdq_s8(v1, v2));
+    acc1 =
+        vaddq_s32(acc1, vreinterpretq_s32_u32(vpaddlq_u16(vpaddlq_u8(diff1))));
 
     v1 = vld1q_s8(pVect1 + 16);
     v2 = vld1q_s8(pVect2 + 16);
-    int8x16_t diff2 = vabdq_s8(v1, v2);
-    acc2 = vaddq_s32(acc2, vpaddlq_u16(vpaddlq_u8(diff2)));
+    uint8x16_t diff2 = vreinterpretq_u8_s8(vabdq_s8(v1, v2));
+    acc2 =
+        vaddq_s32(acc2, vreinterpretq_s32_u32(vpaddlq_u16(vpaddlq_u8(diff2))));
 
     v1 = vld1q_s8(pVect1 + 32);
     v2 = vld1q_s8(pVect2 + 32);
-    int8x16_t diff3 = vabdq_s8(v1, v2);
-    acc3 = vaddq_s32(acc3, vpaddlq_u16(vpaddlq_u8(diff3)));
+    uint8x16_t diff3 = vreinterpretq_u8_s8(vabdq_s8(v1, v2));
+    acc3 =
+        vaddq_s32(acc3, vreinterpretq_s32_u32(vpaddlq_u16(vpaddlq_u8(diff3))));
 
     v1 = vld1q_s8(pVect1 + 48);
     v2 = vld1q_s8(pVect2 + 48);
-    int8x16_t diff4 = vabdq_s8(v1, v2);
-    acc4 = vaddq_s32(acc4, vpaddlq_u16(vpaddlq_u8(diff4)));
+    uint8x16_t diff4 = vreinterpretq_u8_s8(vabdq_s8(v1, v2));
+    acc4 =
+        vaddq_s32(acc4, vreinterpretq_s32_u32(vpaddlq_u16(vpaddlq_u8(diff4))));
 
     pVect1 += 64;
     pVect2 += 64;
@@ -348,8 +354,9 @@ static i32 l1_int8_neon(const void *pVect1v, const void *pVect2v,
   while (pVect1 < pEnd1 - 15) {
     int8x16_t v1 = vld1q_s8(pVect1);
     int8x16_t v2 = vld1q_s8(pVect2);
-    int8x16_t diff = vabdq_s8(v1, v2);
-    acc1 = vaddq_s32(acc1, vpaddlq_u16(vpaddlq_u8(diff)));
+    uint8x16_t diff = vreinterpretq_u8_s8(vabdq_s8(v1, v2));
+    acc1 =
+        vaddq_s32(acc1, vreinterpretq_s32_u32(vpaddlq_u16(vpaddlq_u8(diff))));
     pVect1 += 16;
     pVect2 += 16;
   }