dangernoodle-io / TaipanMiner / 25245916521

90%

push

github

perf(mining): persistent SHA_TEXT zeros across nonces (TA-320b phase A) (#309)

Skip per-nonce writes to SHA_TEXT[9..14] — these slots are 0 in both
the pass1 (block2 tail) and pass2 (digest tail) padding layouts, and
the SHA peripheral preserves them across SHA_START/SHA_CONTINUE
(verified by sha256_hw_verify_text_preserved boot probe, TA-320).

Prime once per job in sha256_hw_pipeline_prep() called from
hw_prepare_job. mbedTLS may grab the SHA peripheral while mining
yields, so re-prime every job (not once at boot).

Measured: tdongle-S3 mining hashrate 220 → 264 kH/s (+20%).
Microbench peripheral ceiling unchanged at 1.63 us/op (306 kH/s).

Bigger gain than estimated: AHB MMIO stores are dominated by bus
turnaround, not raw cycles, so the 12 skipped stores save more time
than naive cycle math suggests.

707 of 852 branches covered (82.98%)

Branch coverage included in aggregate %.

1607 of 1654 relevant lines covered (97.16%)

249583.16 hits per line