dangernoodle-io / TaipanMiner / 26552447225

90%

push

github

perf(mining): fix D0 early-reject byte order — exact 32-bit target filter (#439)

the classic ESP32 (DPORT) hot-loop early-reject compared the raw SHA
register word sha_text[7] against a byte-reversed target word: hw_prepare_job
packed target[28] as the high byte, while sha_text[7] holds canonical H[7]
in big-endian register form. comparing those two scrambled operands made
the filter pass a fixed ~38% of nonces (= target_word0_max/2^32) regardless
of actual difficulty, instead of acting as a real PoW pre-filter. every
passing nonce then paid full digest readback + meets_target.

ground-truth probe on esp32-wroom32 confirmed: passing nonces had true
hashtop e.g. e122ee54 vs true target 00000063 (meets=0) — i.e. the filter
was letting through values ~10^7x over target.

fix: compare bswap32(sha_text[7]) — the true PoW most-significant word, the
chunk meets_target checks first — against target_word0_max packed in the same
true-MSB order (target[31]=MSB). this makes the early-reject an exact top-32-bit
target compare. pass rate at pool diff 0.006 drops from ~65% to ~0.000002%,
eliminating nearly all share_validate calls and accept-path readbacks.

measured on esp32-wroom32 @ pool diff 0.006:
  - before: 233 kH/s
  - after:  446 kH/s  (+91%), 100% share acceptance, pool confirms shares
applies to both the inline kernel and the sha256_hw_dport_per_nonce fallback.
no effect on s3/ahb boards (separate kernel). closes TA-396.

Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>

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