diff options
| -rw-r--r-- | src/board_representation/MIGRATION.md | 89 |
1 files changed, 62 insertions, 27 deletions
diff --git a/src/board_representation/MIGRATION.md b/src/board_representation/MIGRATION.md index 57ca7f9..313c74d 100644 --- a/src/board_representation/MIGRATION.md +++ b/src/board_representation/MIGRATION.md @@ -15,14 +15,15 @@ inline as a note rather than deleted, so a future session designing the separate migration doesn't have to rediscover it from git history. **Status: sections 1-3 (`bbPieces`/`bbPawns`-backed -`_WhoAttacksSquareBB`/`_GetAttacksBB`) and section 6 (the -`GETATTACKS_BITBOARD` toggle, `_GetAttacksBB` now reachable from real -search when enabled) implemented, verified, and committed. Section 4 -items 1-3 done (folded into sections 3/6's own testing); items 4-5 and -sections 5's second half/7 -- the whole-engine curated-suite and -`match_play.py` gates -- not started, now genuinely unblocked since -section 6 exists. Section 8 not started.** -See the per-section status notes below for specifics. +`_WhoAttacksSquareBB`/`_GetAttacksBB`), section 4 items 1-3 and 5, and +section 6 (the `GETATTACKS_BITBOARD` toggle) all implemented, verified, +and committed -- including a whole-engine `sd10` check across all three +curated suites showing zero solve-count regression and +8.38% +aggregate NPS. Only `match_play.py` (section 4 item 6 / section 7) +remains unmet before full retirement of the old mailbox +implementation, deliberately deferred for now. Section 5's whole-engine +item and section 7's remaining items are otherwise done; section 8 not +started.** See the per-section status notes below for specifics. Context: a prior pass of this session added occupancy-bitboard-driven mobility/attack-presence bitboards (`bbPawnAttacks`, `bbMinorAttacks`, @@ -467,16 +468,45 @@ sequence. directly (`generate.c`'s 5 call sites, `searchsup.c:365`), so even value-for-value-identical results can shift which move gets tried first or which lines get pruned/extended, changing node counts and - occasionally search results at the margins. Run all three curated - suites (`ecm_ringers`, `ecm_confident_quick`, `ecm_hard_quick`) at - `sd10` against `head_reference/`, and don't treat a clean - `TestGetAttacks` unit-level pass as sufficient sign-off by itself. - Not started -- requires section 6's toggle to exist first, so there's - something to flip on for the suite runs. + occasionally search results at the margins. **DONE** -- run all + three curated suites at `sd10`, but *not* against the checked-in + `head_reference/` binary: that binary predates this branch's + sections 1-6 by a dozen-plus unrelated commits, so diffing against + it would conflate this change with everything else on the branch. + Instead, isolated the one variable that matters: same commit, built + twice (`gmake clean && gmake -j5 GENETIC=1 PERF_COUNTERS=1 MP=1 + SIXTYFOUR=1` with and without `GETATTACKS_BITBOARD=1`), same + `--cpus 1 --hash 256m`, `sd 10`, `book name /nonexistent.book.bin` + invocation as `head_reference/`'s own convention. Results: + + | suite | solves (asm -> BB) | nodes (asm -> BB) | NPS (asm -> BB) | + |---|---|---|---| + | `ecm_ringers` (11) | 10 -> 10 | 37,611,927 -> 37,909,197 (+0.79%) | 1,178,986 -> 1,270,436 (+7.76%) | + | `ecm_confident_quick` (90) | 83 -> 83 | 429,387,924 -> 430,930,314 (+0.36%) | 1,201,203 -> 1,300,133 (+8.24%) | + | `ecm_hard_quick` (90) | 25 -> 25 | 459,712,049 -> 465,403,216 (+1.24%) | 1,131,433 -> 1,228,991 (+8.63%) | + + **Solve counts bit-identical on all three suites -- zero + regression.** Node counts up slightly (+0.36% to +1.24%), exactly + the "value-identical but not order-identical" effect warned about + above -- `_GetAttacksBB` returns the same attacker *set* as asm + `GetAttacks` but not necessarily the same order, so `SEE()`/move- + ordering tie-breaking can shift which lines get searched first at + the margins. Didn't cost a single solve here. Aggregate (total + nodes / total script time across all three suites, not an + average-of-averages): baseline 926,711,900 nodes / 795.7s = + 1,164,842 nps; bitboard 934,242,727 nodes / 740.0s = 1,262,490 nps + -- **+8.38% overall wall-clock NPS**, comfortably absorbing the + extra node count. Whole-engine result confirms section 3's isolated + cycles/call benchmark wasn't an artifact of measuring the wrong + thing -- the speedup shows up in real search throughput, not just + in the isolated per-call harness. 6. **`match_play.py` gate** (`LOWER95 >= 0.5`) before calling this done -- same reasoning: this is closer to a search-behavior change than a pure eval-magnitude change, so the existing eval-tuning gate criteria - apply. Not started, same section-6 dependency as above. + apply. Not started yet, deliberately deferred -- item 5's zero solve + regression + throughput win is convincing enough on its own for now; + this gate is still required before full retirement (section 7), just + not being chased immediately. ## 5. Microbenchmarking @@ -498,14 +528,17 @@ visible in isolation. Two tiers, both required before wiring in: asm opening, 0.55x middlegame, 0.89x endgame). Consistent win across the spectrum, no red flag. 2. **Whole-engine cycles-per-node / NPS**, not just the isolated call: - `sd`-fixed-depth comparison against `head_reference` on the three - curated suites, using the engine's own self-reported `Searched for N - seconds, M nodes` (not `ps` sampling -- noisy on this shared box per - earlier sessions). This is the check that would have caught the Eval - regression earlier if it had been run before the reader-migration - work proceeded -- don't skip straight from isolated-cycles-looks-good - to production wiring again. Not started -- requires section 6's - toggle to exist first. + `sd`-fixed-depth comparison on the three curated suites, using the + engine's own self-reported node counts/NPS (not `ps` sampling -- + noisy on this shared box per earlier sessions). This is the check + that would have caught the Eval regression earlier if it had been + run before the reader-migration work proceeded -- don't skip + straight from isolated-cycles-looks-good to production wiring + again. **DONE**, folded into section 4 item 5's same-commit A/B run + (see there for the numbers and why `head_reference/` itself wasn't + the comparison point) -- +8.38% aggregate wall-clock NPS across all + three suites, confirming the isolated-cycles-per-call win (section + 3) actually shows up end to end, not just in the harness. ## 6. Dual-support / toggle strategy -- DONE @@ -562,10 +595,12 @@ Only delete `SlowGetAttacks`/asm `GetAttacks` after **all** of: - Isolated cycles/call shows a consistent win across the piece-density spectrum (not just one favorable position). **DONE.** - Whole-engine `sd10` on all three curated suites shows no solve-count - regression vs. `head_reference`. Not started -- needs section 6's - toggle first. -- `match_play.py` gate clears `LOWER95 >= 0.5`. Not started, same - dependency. + regression. **DONE** -- bit-identical solve counts on all three + (10/11, 83/90, 25/90), same-commit asm-vs-`GETATTACKS_BITBOARD=1` A/B + rather than vs. a now-stale `head_reference/` (see section 4 item 5 + for why and the full numbers) -- also +8.38% aggregate wall-clock NPS. +- `match_play.py` gate clears `LOWER95 >= 0.5`. Not started -- + deliberately deferred for now, still required before retirement. - `head_reference/` rebuilt as the new baseline once landed. One easy win to check first, independent of all this: confirm whether |