dangernoodle-io / TaipanMiner / 26673846412

90%

push

github

fix(mining/sha256_hw_ahb): bench us_per_op now per-SHA-op (matches DPORT, fleet-comparable) (#456)

- ahb sha256_hw_bench_pass2 previously returned us_per_op = elapsed_us / iters (per-nonce); dport divides by 3 (per-sha-op). same field, different meaning per-board.
- c3 verified post-#455: khs=378, sha_ops_per_sec=378143, us_per_op=2.64 → sha_ops_per_sec/(khs*1000) ≈ 1 not 3 (would have been silent if not bench-compared with wroom32).
- discovery: ahb uses the midstate optimization — block1 is computed once per job, not per nonce. per-nonce the hot loop does 2 sha ops (pass1: sha_continue block2 tail; pass2: sha_start outer digest), not 3 like dport. ahb now divides by 2 → us_per_op = per-sha-block-op fleet-wide.
- host test exercises both ahb-shaped (ratio ≈ 2) and dport-shaped (ratio ≈ 3) snapshots; invariant now binds the live helpers, not just the json builder.
- closes the bench-semantic half of TA-401 (yield-cadence half deferred to its own measurement pass).
- api shape unchanged for callers reading khs (still nonces/sec/1000); us_per_op numeric value changes on ahb boards (now ~1.32 µs/op on c3, was 2.64 µs/nonce); sha_kops_per_sec numeric value changes on ahb boards (now ~756k on c3, was 378k).

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

2593 of 3258 branches covered (79.59%)

Branch coverage included in aggregate %.

5247 of 5422 relevant lines covered (96.77%)

152316.02 hits per line