dangernoodle-io / TaipanMiner / 26551484675

90%

push

github

perf(mining): share_validate fast-path + honest SHA microbench (#438)

* perf(mining): share_validate fast-path + honest SHA microbench

at low pool diff the D0/DPORT hw pre-filter passes ~65% of nonces into
share_validate, where redundant is_target_valid + diff floor checks
(both job-invariant, already validated pre-loop at mining.c:914) burned
~330 cyc per HASH_CHECK. inlining meets_target into the hot loop and
deferring share_validate to actual candidates drops per-HASH_CHECK
work from 369→37 cycles. measured recovery on esp32-wroom32:
  - low diff 0.006:  233 -> 307 kH/s (+32%)
  - high diff 0.01:  286 -> 307 kH/s (+7%)
tdongle-s3 unchanged (s3 ahb kernel filter is ~10000x tighter; ~0.001%
hashcheck rate, so the fast-path saves nothing there).

microbench fix (TA-395): previous formula `khs = 500/us_per_op` assumed
2 sha ops/nonce — correct for s3 (midstate preload) but wrong for d0
(no midstate; 3 ops/nonce) and ignored fill-during-compute overlap on
both. replace with real-kernel timing:
  - d0: call sha256_hw_dport_kernel(target=0) 1000x, measure cycles
  - s3: reuse existing sha256_hw_profile_hotloop output, wire it into
    mining_set_sha_microbench
result: d0 ceiling 418 -> 548 kH/s (honest reject-path bound — closes
the gap to silicon when TA-396 lands); s3 ceiling 304 -> 317 kH/s
(slight bump from including pipeline overlap, now eff ~97% matches
observed silicon).

diag instrumentation kept behind MINING_HOTLOOP_DIAG (default off,
compiles out): kernel_cyc_avg / hashcheck / validate_cyc_avg probes
proved invaluable for the bisect, leaving them gated so the next
regression investigation can flip one flag.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* chore(mining): log pool acceptance ratio in hw: line

surfaces accepted/rejected counts + percentage alongside hashrate so
the share_validate fast-path's correctness can be verified at a glance
in /api/logs. if fast-path silently drops real shares the ratio falls
below historical ... (continued)

2549 of 3217 branches covered (79.24%)

Branch coverage included in aggregate %.

5158 of 5332 relevant lines covered (96.74%)

154876.85 hits per line