dangernoodle-io / TaipanMiner / 26695115133

90%

push

github

fix(mining): bench helpers measure full nonce work, not single SHA op (#459)

- ahb sha256_hw_bench_pass2 was measuring 1 SHA_CONTINUE per iter but dividing us_per_op by 2 and treating iters as nonces in khs calc — net khs ~2x too high (tdongle bench 455 vs live mining 310, s2 bench 436 vs live 266)
- now: loop body calls sha256_hw_mine_nonce (the actual production hot-loop function) per iter → both passes + overlap optimization, iters = nonces, /2 normalization on us_per_op is now correct
- dport bench was already calling full kernel; only cold-cache variance affected the wroom32 ceiling — bump boot microbench iters 1000 → 5000 (same for ahb profile_hotloop) to reduce noise
- restores historical eff% (tdongle ≈ 99-101%, wroom32 ≈ 65-70% per CLAUDE.md)
- ta-395 was the upstream concern; this is the final-mile correction
- bench loop body now models actual mining work (not a synthetic micro-pattern) so future optimizations to the hot loop reflect in the bench

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

2594 of 3258 branches covered (79.62%)

Branch coverage included in aggregate %.

5247 of 5422 relevant lines covered (96.77%)

152315.93 hits per line