2026-09-10. The preceding pinned-transfer version measured 95.0–95.4 ms unsanitized / 115.9 ms UBSan. Warm CPU samples attributed 8.15% to skinning.
SAM3D_SIMD_SKINNING=1 processes four independent influences with baseline
SSE2 on x86. It preserves both quaternion normalizations, each multiply/add/
cross-product's order, and serial scatter-add order into mesh vertices. Duplicate
vertex IDs are not reduced in parallel. There is no matrix approximation,
fast-math, FMA contraction, influence pruning, validation bypass or model change.
The scalar path handles tails, diagnostic operation capture and other CPUs.
Without the flag the previous path is retained. No C ABI or GGML patch changed.
- All 48 CPU ASan/UBSan/LSan and Vulkan-UBSan normal tests pass.
- The existing original skinning fixture still checks 12 boundaries. It now also exercises the SIMD final output. Ninety-six added deterministic cases compare scalar, SIMD and captured outputs bit for bit, including one/two batches, varied rotations/scales/translations, shuffled and duplicate vertex indices, zero weights, changing influence counts and every short tail.
- All 25 BF16 outputs in each accepted timing/profile run are byte-identical to the previous baseline. The normal public F32 path also matches all 25 previous outputs; this is important because operation capture deliberately remains scalar and cannot alone validate the fast path.
- All 529 F32 captured tensors retain their previous hash. Frozen original checks pass all 646 assertions, not replaced with a fast-versus-slow comparison.
- Both final BF16 policies pass with the same reported non-blocking dancer hand-logit difference. All 112 Python tests and Go race tests pass.
All runs have five warmups and twenty timed alternating official images, resident weights and complete host outputs. BF16 encoder/F32 decoder and all six pose/MHR feedback layers remain enabled.
| Run | SIMD on | Same-binary off |
|---|---|---|
| First | 94.165 ms | 95.573 ms |
| Repeat | 93.803 ms | 95.294 ms |
The repeat's p95 is 94.597 ms, range 93.539–95.853 ms. UBSan-O2 measures 114.150 ms (p95 114.869 ms). CPU sampling measures 94.159 ms; skinning's sample share falls to 3.52%, with fence polling 44.51%, copying 5.44% and image preparation 4.93%. These sample fractions are not additive wall-clock stages. GPU arithmetic and transferred bytes are unchanged. The 80–85 ms goal stays open; GPU work now offers more potential than further skinning work.
Local ignored evidence: generated/diagnostics/bf16-skin-simd-*,
generated/fixtures/bf16-skin-simd-full-f32-v1/. Heavy jobs are serialized,
memory bounded and retain the 10 GiB reserve. The GGML submodule is pristine;
no Nix derivation rebuilding is involved.
The existing demo was updated in place with the UBSan build and saved history
retained. The server selects SIMD skinning only for its Vulkan BF16 mode;
inherited conflicting precision switches are stripped. Real Chrome 151 QA
passes upload, cancellation/recovery, warm reuse, original overlay, orbit
controls, history/mobile scrolling and exact GLB export/reload. The front
overlay was visually inspected; body and arm geometry are unchanged. Native
and GLB hashes remain identical. Cold/warm browser jobs took 4.349 / 0.169 s
including export, not benchmark medians. Evidence:
generated/diagnostics/bf16-skin-simd-demo-qa-v1/.