Spike note — ARM NEON packed GEMM (W2.T04)

Field	Value
Status	DELIVERED — see ../41_PORTABLE_SIMD_KERNEL.md
Date	2026-07-25 (note) · delivered with portable SIMD
Blocker	~~No aarch64 CI runner~~ — resolved via GitHub ubuntu-24.04-arm + macos-latest
Acceptance	Met: vcntq_u8/vpadalq_u8 NEON path + aarch64/macOS CI asserting err=0

Intent (historical)

Edge phones / Apple Silicon are the honest BNN deployment story. At spike-write time the matrix was:

• Windows x64 — MSVC OpenMP DLL (primary lab path)
• Linux x64 — GCC .so (CI hard gate as of v0.3)
• macOS / ARM — NumPy correctness fallback only ← superseded

Delivered (current)

1. #ifdef / runtime NEON path in bnn/kernels/binary_gemm.c using vcntq_u8 / vpadalq_u8 (128-bit words).
2. Scalar / __builtin_popcountll fallback kept for correctness.
3. compile_native on Darwin + aarch64 Linux (Clang/GCC -O3 -shared -fPIC, OpenMP when present).
4. CI: portability job on ubuntu-24.04-arm and macos-latest (plus forced BNN_KERNEL=scalar).
5. Dual-metric bench only — no theory-as-latency claims.

Interim acceptance (met → closed)

• Documented in docs/COMPATIBILITY_MATRIX.md (native NEON on arm64)
• Tests pass via native and NumPy path
• ROADMAP W2.T04 flipped to [x]

Non-goals

• Stock phone NPU 1-bit (vendor INT8/INT4) — see docs/20
• Claiming NEON speedups without measured wall-clock on target silicon