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+ CIBNN_KERNEL=scalarrerun - 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).