threat-inputs vs master diff
ELO 20.76 +- 15.24 (95%) SPRT 60.0+0.60s Threads=1 Hash=128MB LLR 0.49 (-2.94, 2.94) [0.00, 3.00] GAMES N: 486 W: 140 L: 111 D: 235 PENTA [0, 45, 124, 74, 0]
NN: Faster threat delta generation The bitboards a threat lookup needs (occupancy, victim masks per attacker type, and the piece code of every square as four bit planes) are computed once per position. Records are written through uint8_t, which may alias the position, so reading it directly reloaded its bitboards after every record. The changed squares come from the planes. Sliders are looked up only through squares whose occupancy changed: a square occupied in both positions stops the ray in both. speedup.py, MinGW bmi2 with eval-ti-2-120B-permuted embedded, depth 12, 20 rounds: 1,835,108 vs 1,814,096 nps, +1.173% [+0.640%, +1.706%]. No functional change. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
LLR History
| Author | Witek |
| Creation | 1795030920 |
| Dev Source | https://github.com/Witek902/Caissa |
| Dev Sha | c5d8d7727b0a2842efc66f161b1555c13cd4509b |
| Dev Branch | threat-inputs |
| Dev Bench | 6386874 |
| Dev Network | eval-ti-2-120B-permuted |
| Dev Options | Threads=1 Hash=128 |
| Dev Time | 60.0+0.60 |
| Base Source | https://github.com/Witek902/Caissa |
| Base Sha | 0ec07379943a3fc1723a276150c537d5be12d3d6 |
| Base Branch | master |
| Base Bench | 6262536 |
| Base Network | |
| Base Options | Threads=1 Hash=128 |
| Base Time | 60.0+0.60 |
| Opening Book | UHO_Lichess_4852_v1.epd |
| Upload PGNs | FALSE |
| Syzygy WDL | DISABLED |
| Syzygy ADJ. | OPTIONAL |
| Win ADJ. | movecount=3 score=600 |
| Draw ADJ. | movenumber=40 movecount=8 score=3 |
| Scaling | BASE 1000000 NPS |
| # | User | Games | Penta | Pairs | Time | Crash |
|---|