Skip to content

Moonshot note — AVX-512 VPOPCNTDQ (W2.T05)

Field Value
Status DELIVERED — see ../41_PORTABLE_SIMD_KERNEL.md
Date 2026-07-25 (note) · delivered with portable SIMD
Blocker ~~Needs dispatch harness~~ — resolved: runtime cpuid+xgetbv dispatch, AVX-512 never required
Acceptance Met: runtime dispatch, AVX2 + scalar fallbacks, cross-ISA err=0 gate

Intent (historical)

VPOPCNTDQ can accelerate packed popcount on Ice Lake / Zen4-class CPUs. This was framed as a perf moonshot, not a correctness gate (NumPy / scalar / __popcnt64 already err=0).

Delivered (current)

  • Runtime dispatch: AVX-512 VPOPCNTDQ → AVX2 → NEON → scalar in one portable build
  • No -march=native; AVX-512 is used when present, never required to install/build/run
  • Cross-ISA err=0 in tests/test_native_gemm.py + CI BNN_KERNEL=scalar rerun
  • ROADMAP W2.T05 flipped to [x]

Explicit non-claim

Do not advertise “AVX-512 32×” without published wall-clock on listed shapes (docs/BENCH_SHAPES.md + docs/FAIR_EVAL_PROTOCOL.md).