typhoon.git/src/nnueeval.c, branch master

NNUE: re-fit raw-to-score calibration, NNU4 weight format, super-lazy default-on, lazy probe measured net-negative

2026-10-21T01:37:20Z

Fits an affine (a + b*raw) correction into every exported .nnue file (export_nnue_weights.py's CalibrateScoreMapping) so the engine's consumed centipawn output tracks Stockfish's real score scale rather than the net's nominal raw WDL-sigmoid output -- training variance alone shifts this mapping across retrains even with an identical recipe, so it's re-fit and baked in per export rather than assumed constant. Weight format bumped to "NNU4" to carry the two new floats; older NNU1/2/3 files are no longer readable. Super-lazy (material-only) pre-filter promoted to default-on (g_fUseNnueSuperLazy = TRUE) with a recalibrated per-material-bucket margin table (measured against the recalibrated net, 16-shard cross_sectional_200k.ep_ sample) -- real, measured nps win across all three curated suites. The separate lazy-probe gate (accumulator-based early exit) is measured net-negative on real search traffic at this net's probe-vs-forward latency ratio and stays off by default (g_fUseNnueLazyProbe = FALSE); code kept live since margin tuning could still flip this, per docs/NNUE.md. Also: direct per-call latency instrumentation for NnueProbeScore/ NnueForward (root.c, EVAL_TIME build) to make the lazy-probe breakeven math measurable directly instead of inferred; QSearch futility margins given real-data-derived NNUE_FULL/NNUE_LAZY starting points. Co-Authored-By: Claude Sonnet 5

Wire NNUE eval into search, build the training pipeline, and add a fast

2026-10-20T16:26:44Z

move/eval-agreement benchmark against Stockfish (tests/lop) Implements the architecture decisions in src/docs/NNUE.md: HalfKAv2_hm feature set, int16/int8 quantization, per-ply accumulator backup for king moves, and a lazy-eval-ready integration into search.c/eval.c. Adds the full annotate_pgns/ training pipeline (extraction, K-fit/ labeling, quiet filtering, tensor building, training, export) plus an end-to-end orchestrator (run_nnue_pipeline.py) and validation tooling (sanity_check_net.py, compare_vs_stockfish_net.py). Adds tests/lop: a 10k-position list-of-positions benchmark comparing engine eval rank and move choice against pre-computed Stockfish/lc0 references, much cheaper than a full self-play match for day-to-day progress checks. First sd10-matched self-play result against classic HCE: LLR +2.28 (H1=+2.94), -72=119+169 over 360 games, confirming the net is "smarter per node" at matched depth. A real-time (--st 4) match is currently weaker (candidate loses ~3 plies of depth to a ~3.3-3.7x nps gap against an unrecalibrated search), consistent with NNUE.md's adoption-criterion section -- pruning/reduction constants still need retuning for the new eval's score scale before a fair real-time comparison is possible. Co-Authored-By: Claude Sonnet 5

Wire NNUE eval into real search path; fix critical perf regression

2026-10-15T23:26:09Z

Load-and-call infra: export_nnue_weights.py-produced weight files can now be loaded at runtime (nnueeval.c's NnueLoadWeights, "nnue load" command, --nnueweights CLI flag for scripted match-play), and the FC head (NnueForward) runs the real forward pass against the live accumulator. Real integration: Eval() now short-circuits to NnueForward when g_fUseNnueEval && g_fNnueWeightsLoaded, so ./typhoon can actually play on the NNUE eval -- ./eval_tune/test_vs_head.sh works unmodified to compare it against the old HCE. ReInitializeSearcherContext (root.c) gained a defensive accumulator recompute hook covering setboard's downstream effect on the next go/ponder; it only fires when both kings are validly placed, since testrecogn.c's synthetic positions don't always have one at the moment this is called. Critical fix: the NNUE accumulator maintenance added to move.c's 8 primitives (and MakeMove/UnmakeMove's king-move backup/restore) was running unconditionally, on the mistaken assumption this was cheap. It cost ~50% of engine throughput even with NNUE off (673k nps measured vs the expected 1.4-1.5M baseline). Gated all of it behind g_fUseNnueEval; classic HCE is back to ~1.52M nps, NNUE eval runs at a fair ~409k nps (expected for an unquantized float32 forward pass). This gating also exposed that testnnue.c's TestNnueAccumulator never set g_fUseNnueEval, so its own direct NnueAccumRecompute calls raced ahead of now-gated-off incremental maintenance; fixed by having the test set/restore the flag around its run. board.c: added en passant legality check to IsBoardLegal (rank vs side-to-move, origin-square occupancy) -- a real, previously-latent bug surfaced while building the NNUE test fixtures, unrelated to NNUE itself. Co-Authored-By: Claude Sonnet 5