Stockfish Testing Queue

Pending - 0 tests 0.0 hrs

None

Active - 0 tests

Finished - 40 tests

18-09-10 omo ghi_cap_inv diff
LLR: -2.43 (-2.94,2.94) [-3.00,1.00]
Total: 9647 W: 2029 L: 2183 D: 5435
sprt @ 10+0.1 th 1 Take 3 (experimental), test for no regression at low throughput.
18-09-10 omo ghi_cap diff
LLR: -1.03 (-2.94,2.94) [0.00,5.00]
Total: 729 W: 149 L: 192 D: 388
sprt @ 10+0.1 th 1 Take 2 (triple fixed)
18-09-10 omo ghi_cap diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 1052 W: 192 L: 321 D: 539
sprt @ 10+0.1 th 1 Take 2: Try capping the counter instead of right-shifting.
18-09-10 omo ghi diff
LLR: -2.94 (-2.94,2.94) [0.00,5.00]
Total: 9124 W: 1910 L: 1996 D: 5218
sprt @ 10+0.1 th 1 My take on Hanamuke's GHI solution idea. My idea is to reduce impact on hash table by right-shifting the 50-move rule counter (reducing its resolution). I'm not sure if it makes any sense, but from local tests it seems like possible ELO gain, so let's see...
18-09-02 omo 4aa091cf4486 diff
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 22905 W: 4961 L: 4842 D: 13102
sprt @ 10+0.1 th 1 Regression test for "Refactor pure static eval code", since there are some doubts. It does produce different machine code, so I guess it warrants a test for no slowdown.
18-06-26 omo ocbSpeedup diff
LLR: -2.95 (-2.94,2.94) [0.00,4.00]
Total: 29259 W: 6489 L: 6537 D: 16233
sprt @ 10+0.1 th 1 Small speed-up of Position::opposite_bishops() No functional change.
18-06-10 omo stateInfoSimpl diff
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 78820 W: 15775 L: 15760 D: 47285
sprt @ 10+0.1 th 1 Simplify do/undo_move by moving PSQ score from StateInfo to Position and updating it in the put/remove/move_piece helpers. Test for no speed regression.
18-06-09 omo npmToPosition diff
LLR: -2.96 (-2.94,2.94) [0.00,4.00]
Total: 65512 W: 13063 L: 12983 D: 39466
sprt @ 10+0.1 th 1 This seems to be a small speedup on my machine.
18-06-03 omo doMoveOpt diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 13514 W: 2612 L: 2678 D: 8224
sprt @ 10+0.1 th 1 Test with copying all keys failed quite strongly, so let's try going the other way around -- *not* copying also pawnKey and materialKey. This time test for ELO gain (speedup).
18-06-02 omo siCopyKey diff
LLR: -2.96 (-2.94,2.94) [-3.00,1.00]
Total: 34654 W: 6771 L: 6995 D: 20888
sprt @ 10+0.1 th 1 Test this slight non-functional simplification for no regression (in speed).
18-05-27 omo pawn-hash-mini diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 6497 W: 1246 L: 1344 D: 3907
sprt @ 10+0.1 th 1 An experiment to see if compressing pawn hash entries (to half the original size) and doubling their count is an improvement. Half throughput.
18-05-27 omo givesCheckCastling diff
LLR: -2.94 (-2.94,2.94) [0.00,4.00]
Total: 53814 W: 10843 L: 10804 D: 32167
sprt @ 10+0.1 th 1 Specialize pos.gives_check() for castlings. Test this non-functional change for speed-up.
18-05-08 omo simplifyBB diff
LLR: -2.95 (-2.94,2.94) [-3.00,1.00]
Total: 117308 W: 23466 L: 23844 D: 69998
sprt @ 10+0.1 th 1 Remove SquareBB, FileBB, and RankBB and replace them with direct computation.
18-05-08 omo simplifyBB^ diff
LLR: -2.95 (-2.94,2.94) [-3.00,1.00]
Total: 156904 W: 31647 L: 32100 D: 93157
sprt @ 10+0.1 th 1 Drop only SquareBB and FileBB (rank_bb requires one more instruction).
18-05-04 omo kingRingPrecompute diff
LLR: -2.95 (-2.94,2.94) [0.00,4.00]
Total: 24054 W: 4841 L: 4912 D: 14301
sprt @ 10+0.1 th 1 Precompute king ring in bitboard.cpp
18-04-17 omo tbppSpeedup diff
LLR: -2.96 (-2.94,2.94) [0.00,4.00]
Total: 16717 W: 3344 L: 3443 D: 9930
sprt @ 10+0.1 th 1 It is useless to compute ThreatByPawnPush if we have no pawns. See if this is a reasonable speedup.
18-03-24 omo dynCon_revert diff
ELO: -11.43 +-4.1 (95%) LOS: 0.0%
Total: 20000 W: 7085 L: 7743 D: 5172
20000 @ 5+0.05 th 3 Skill Level=17, 3 threads: revert dynamic contempt vs. dynCon_prevScore only
18-03-24 omo dynCon_revert diff
ELO: 10.98 +-4.1 (95%) LOS: 100.0%
Total: 20000 W: 7673 L: 7041 D: 5286
20000 @ 5+0.05 th 3 Skill Level=17, 3 threads: revert dynamic contempt vs. master
18-03-24 omo dynCon_revert diff
ELO: -26.42 +-4.1 (95%) LOS: 0.0%
Total: 20000 W: 6661 L: 8179 D: 5160
20000 @ 5+0.05 th 3 Skill Level=17, 3 threads: revert dynamic contempt vs. dynCon_prevScore + dynCon_perThread_ng
18-03-24 omo dynCon_revert diff
ELO: -30.27 +-4.2 (95%) LOS: 0.0%
Total: 20000 W: 6562 L: 8300 D: 5138
20000 @ 10+0.1 th 1 Skill Level=17, 1 thread: revert dynamic contempt vs. dynCon_prevScore
18-03-24 omo dynCon_revert diff
ELO: 30.92 +-4.1 (95%) LOS: 100.0%
Total: 20000 W: 8298 L: 6523 D: 5179
20000 @ 10+0.1 th 1 Skill Level=17, 1 thread: revert dynamic contempt vs. master
18-03-20 omo dynCon_perThread diff
LLR: -2.97 (-2.94,2.94) [0.00,5.00]
Total: 41396 W: 7127 L: 7082 D: 27187
sprt @ 5+0.05 th 16 Try per-thread contempt with many (16) threads. Half throughput.
18-03-23 omo dynCon_perThread_ng diff
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 65575 W: 12786 L: 12745 D: 40044
sprt @ 5+0.05 th 5 Now verify for no regression with MultiPV=1. Test against the dynCon_prevScore patch.
18-03-23 omo dynCon_perThread_ng diff
LLR: 2.95 (-2.94,2.94) [0.00,5.00]
Total: 1279 W: 549 L: 380 D: 350
sprt @ 5+0.05 th 5 Test also both patches together to ensure there is no significant regression in the first one (for multi-thread + Skill Level=17).
18-03-23 omo dynCon_perThread_ng diff
LLR: 2.96 (-2.94,2.94) [0.00,5.00]
Total: 3498 W: 1316 L: 1135 D: 1047
sprt @ 5+0.05 th 5 Check if per-thread contempt is better for multi-threaded MultiPV search (Skill Level=17). Test against the dynCon_prevScore patch.
18-03-19 omo dynCon_prevScore_minima diff
LLR: 2.96 (-2.94,2.94) [0.00,5.00]
Total: 2399 W: 861 L: 694 D: 844
sprt @ 240+2.4 th 1 Use prevScore of correct root move for dynamic contempt. Test at VLTC in case of empty framework (I'm worried about the low draw rate). Hopefully this passes quickly like the other tests.
18-03-21 omo dynCon_perThread' diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 27164 W: 4974 L: 4983 D: 17207
sprt @ 5+0.05 th 8 With only a few 16-core machines now, I'd like to try an 8-thread test and see if that's enough to see an improvement. Also disable thread binding as per CoffeeOne's suggestion.
18-03-16 omo c5f6bd517c68 diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 26745 W: 8726 L: 8709 D: 9310
sprt @ 60+0.6 th 1 Test a known ELO gain with Skill Level=17 to see if Skill Level tests are actually usable for proving improvement.
18-03-16 omo dynCon_prevScore_minima diff
LLR: 2.95 (-2.94,2.94) [0.00,5.00]
Total: 747 W: 333 L: 175 D: 239
sprt @ 60+0.6 th 1 Use prevScore of correct root move for dynamic contempt. Test with Skill Level=17 (implicitly MultiPV=4). Just LTC, since MutliPV slows down the search. (Turns out quotes around Skill Level do not work, drop them.)
18-03-16 omo dynCon_prevScore_minima diff
LLR: -1.40 (-2.94,2.94) [0.00,5.00]
Total: 33655 W: 5113 L: 5046 D: 23496
sprt @ 60+0.6 th 1 Use prevScore of correct root move for dynamic contempt. Test with Skill Level=17 (implicitly MultiPV=4). Just LTC, since MutliPV slows down the search.
18-03-15 omo dynCon_prevScore_minima diff
LLR: 2.95 (-2.94,2.94) [0.00,5.00]
Total: 2390 W: 874 L: 706 D: 810
sprt @ 60+0.6 th 1 Use prevScore of correct PV move for dynamic contempt. Non-functional for MultiPV=1. Test if this is better with MultiPV=3. (LTC, clean version)
18-03-15 omo dynCon_prevScore diff
LLR: 2.95 (-2.94,2.94) [0.00,5.00]
Total: 2657 W: 1079 L: 898 D: 680
sprt @ 10+0.1 th 1 Use prevScore of correct PV move for dynamic contempt. Non-functional on 1 thread and MultiPV=1. Test if this is better with MultiPV=3.
18-03-14 omo dynCon_perThread_ng diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 82689 W: 16439 L: 16190 D: 50060
sprt @ 5+0.05 th 5 Retry per-thread dynamic contempt over the new atan formula. (Fixed compilation error)
18-03-01 omo symmetricContempt_linea diff
LLR: -2.95 (-2.94,2.94) [-3.00,1.00]
Total: 118173 W: 24064 L: 24447 D: 69662
sprt @ 10+0.1 th 1 Take2: Contempt 10 cp +/- 19 (vs Contempt=10). I messed something up, resubmitting...
18-02-28 omo symmetricContempt_linea diff
LLR: -2.95 (-2.94,2.94) [-3.00,1.00]
Total: 48317 W: 9818 L: 10070 D: 28429
sprt @ 10+0.1 th 1 Try to simplify Stefano's logarithmic contempt.
18-02-28 omo symmetricContempt_linea diff
LLR: -0.06 (-2.94,2.94) [-3.00,1.00]
Total: 891 W: 180 L: 185 D: 526
sprt @ 10+0.1 th 1 Try to simplify Stefano's logarithmic contempt.
18-02-06 omo dynamic_perThread diff
ELO: 0.64 +-2.2 (95%) LOS: 71.4%
Total: 40000 W: 8653 L: 8579 D: 22768
40000 @ 5+0.05 th 5 Try per-thread dynamic contempt (on 5 threads)
18-02-12 omo dynamic_random diff
LLR: -1.26 (-2.94,2.94) [-3.00,1.00]
Total: 524 W: 78 L: 148 D: 298
sprt @ 10+0.1 th 1 An experiment to see if dynamic contempt is equivalent to random contempt offset. Low throughput.
18-02-09 omo dynamic_perThread_v2 diff
LLR: -2.97 (-2.94,2.94) [0.00,5.00]
Total: 17368 W: 2798 L: 2853 D: 11717
sprt @ 30+0.3 th 5 Let's see if per-thread contempt is better than the new master at SMP LTC.
18-02-10 omo unordered_map diff
LLR: -2.95 (-2.94,2.94) [0.00,4.00]
Total: 9647 W: 2074 L: 2200 D: 5373
sprt @ 10+0.1 th 1 Use unordered_map for endgame bases. This is probably not going to make a difference, but I'd like to try it anyway.