ternlogq tail: padded zmm vs zmm→ymm→xmm descent, measured (AVX-512) - #311

Merged

AdaWorldAPI merged 8 commits into

claude/c64-6502-falsifier-shztkk

Sep 16, 2026

Commits on Sep 16, 2026

ternlogq tail: measure padded zmm vs zmm→ymm→xmm descent (AVX-512 only)
claude
committed
9ddb531
ternlogq sparse re-apply: chunk-skip zmm/ymm/xmm vs the gather, measured
claude
committed
3124eeb
sgemm block-stop probe: an M-tail R×16 ukernel is measured inert
claude
committed
3c5915b
clippy --tests: regroup the masking-ops test seed literal (unusual_byte_groupings)
claude
committed
07e14e3
blackboard (13): the three tail probes, full tables — descent 5-8x, sparse re-apply no, GEMM block-stop inert
claude
committed
23b4e87
ternlogq probes: gate the active main on avx512vl (+dq for vptestmq), not avx512f alone
claude
committed
84be8d0
ternlogq_sparse_reapply_probe: the cfg gate itself now requires avx512vl + avx512dq
claude
committed
6986d48
probes: SAFETY comment on every unsafe block; fallback messages name every gated feature
claude
committed
f164f53