stockfish
13 months agoSchedule threads fairly under valgrind
Joost VandeVondele [Wed, 30 Sep 2020 19:22:36 +0000 (21:22 +0200)]
Schedule threads fairly under valgrind

fixes a rare case that can cause CI to fail when running multithreaded under valgrind.

closes https://github.com/official-stockfish/Stockfish/pull/3165

No functional change.

14 months agoInclude pawns in NNUE scaling
Stéphane Nicolet [Mon, 28 Sep 2020 20:32:55 +0000 (22:32 +0200)]
Include pawns in NNUE scaling

We now include the total pawn count in the scaling factor for the output
of the NNUE evaluation network. This should have the effect of trying to
keep more pawns when SF has the advantage, but exchange them when she
is defending.

Thanks to Alexander Pagel (Lolligerhans) for the idea of using the
value of pawns to ease the comparison with the rest of the material
estimation.

Passed STC:
LLR: 2.93 (-2.94,2.94) {-0.25,1.25}
Total: 15072 W: 1700 L: 1539 D: 11833
Ptnml(0-2): 65, 1202, 4845, 1355, 69
https://tests.stockfishchess.org/tests/view/5f7235a63b22d6afa50699b3

Passed LTC:
LLR: 2.93 (-2.94,2.94) {0.25,1.25}
Total: 25880 W: 1270 L: 1124 D: 23486
Ptnml(0-2): 23, 980, 10788, 1126, 23
https://tests.stockfishchess.org/tests/view/5f723b483b22d6afa5069a99

closes https://github.com/official-stockfish/Stockfish/pull/3164

Bench: 3776081

14 months agoTweak reduction formula.
Stefan Geschwentner [Mon, 28 Sep 2020 20:21:14 +0000 (22:21 +0200)]
Tweak reduction formula.

Replace log(i) with log(i + 0.25 * log(i)). This increases especially for low values the reductions. But for bigger values there are nearly no changes.

STC:
LLR: 2.94 (-2.94,2.94) {-0.25,1.25}
Total: 49640 W: 5505 L: 5289 D: 38846
Ptnml(0-2): 270, 4074, 15924, 4274, 278
https://tests.stockfishchess.org/tests/view/5f71f04d3b22d6afa5069478

LTC:
LLR: 2.94 (-2.94,2.94) {0.25,1.25}
Total: 43856 W: 2209 L: 2021 D: 39626
Ptnml(0-2): 32, 1776, 18128, 1956, 36
https://tests.stockfishchess.org/tests/view/5f7232ee3b22d6afa50699a2

closes https://github.com/official-stockfish/Stockfish/pull/3163

Bench: 3555769

14 months agoUpdate default net to nn-baeb9ef2d183.nnue
SFisGOD [Mon, 28 Sep 2020 18:24:26 +0000 (02:24 +0800)]
Update default net to nn-baeb9ef2d183.nnue

Further optimization of Sergio's nn-03744f8d56d8.nnue
This patch is the result of collaboration with Joost VandeVondele.

STC:
LLR: 2.96 (-2.94,2.94) {-0.25,1.25}
Total: 37000 W: 4145 L: 3947 D: 28908
Ptnml(0-2): 191, 3016, 11912, 3166, 215
https://tests.stockfishchess.org/tests/view/5f71e7983b22d6afa5069475

LTC:
LLR: 2.96 (-2.94,2.94) {0.25,1.25}
Total: 60224 W: 2992 L: 2769 D: 54463
Ptnml(0-2): 48, 2420, 24956, 2637, 51
https://tests.stockfishchess.org/tests/view/5f722bb83b22d6afa506998f

closes https://github.com/official-stockfish/Stockfish/pull/3161

Bench: 3720073

14 months agoTweaking Mobility and Safe Check
FauziAkram [Mon, 28 Sep 2020 19:09:43 +0000 (22:09 +0300)]
Tweaking Mobility and Safe Check

Passed STC:
https://tests.stockfishchess.org/tests/view/5f70d86d3b22d6afa50693b9
LLR: 2.94 (-2.94,2.94) {-0.25,1.25}
Total: 100368 W: 20323 L: 19914 D: 60131
Ptnml(0-2): 1927, 11641, 22605, 12118, 1893

Passed LTC:
https://tests.stockfishchess.org/tests/view/5f71bb553b22d6afa5069457
LLR: 2.94 (-2.94,2.94) {0.25,1.25}
Total: 77648 W: 10613 L: 10181 D: 56854
Ptnml(0-2): 634, 7280, 22594, 7652, 664

closes https://github.com/official-stockfish/Stockfish/pull/3160

Bench: 3861984

14 months agoAdjust null move pruning constants
Vizvezdenec [Mon, 28 Sep 2020 17:20:06 +0000 (20:20 +0300)]
Adjust null move pruning constants

Idea is that division by fraction of 2 is slightly faster than by other numbers so parameters are adjusted in a way that division in null move pruning depth reduction features dividing by 256 instead of dividing by 213.
Other than this patch is almost non-functional - difference starts to exist by depth 133.

passed STC
https://tests.stockfishchess.org/tests/view/5f70dd943b22d6afa50693c5
LLR: 2.95 (-2.94,2.94) {-0.25,1.25}
Total: 57048 W: 6616 L: 6392 D: 44040
Ptnml(0-2): 304, 4583, 18531, 4797, 309

passed LTC
https://tests.stockfishchess.org/tests/view/5f7180db3b22d6afa506941f
LLR: 2.95 (-2.94,2.94) {0.25,1.25}
Total: 45960 W: 2419 L: 2229 D: 41312
Ptnml(0-2): 43, 1779, 19137, 1987, 34

closes https://github.com/official-stockfish/Stockfish/pull/3159

bench 3789924

14 months agoUpdate default net to nn-04a843f8932e.nnue
Joost VandeVondele [Sat, 26 Sep 2020 15:47:52 +0000 (17:47 +0200)]
Update default net to nn-04a843f8932e.nnue

an optimization of Sergio's nn-03744f8d56d8.nnue tuning the output layer (33 parameters) on game play.

WIP code to make layer parameters tunable is https://github.com/vondele/Stockfish/tree/optionOutput
Optimization itself is using https://github.com/vondele/nevergrad4sf
Writing of the modified net using WIP code based on the learner code https://github.com/vondele/Stockfish/tree/evalWrite

Most parameters in the output layer are changed only little (~5 for int8_t).

passed STC:
https://tests.stockfishchess.org/tests/view/5f716f6b3b22d6afa506941a
LLR: 2.94 (-2.94,2.94) {-0.25,1.25}
Total: 15488 W: 1859 L: 1689 D: 11940
Ptnml(0-2): 79, 1260, 4917, 1388, 100

passed LTC:
https://tests.stockfishchess.org/tests/view/5f71908e3b22d6afa506942e
LLR: 2.93 (-2.94,2.94) {0.25,1.25}
Total: 8728 W: 518 L: 400 D: 7810
Ptnml(0-2): 7, 338, 3556, 456, 7

closes https://github.com/official-stockfish/Stockfish/pull/3158

Bench: 3789924

14 months agoUse incremental updates more often
noobpwnftw [Sun, 27 Sep 2020 18:29:21 +0000 (02:29 +0800)]
Use incremental updates more often

Use incremental updates for accumulators for up to 2 plies.
Do not copy accumulator. About 2% speedup.

Passed STC:
LLR: 2.95 (-2.94,2.94) {-0.25,1.25}
Total: 21752 W: 2583 L: 2403 D: 16766
Ptnml(0-2): 128, 1761, 6923, 1931, 133
https://tests.stockfishchess.org/tests/view/5f7150cf3b22d6afa5069412

closes https://github.com/official-stockfish/Stockfish/pull/3157

No functional change

14 months agoTweak nnue scaling to keep more material
Stéphane Nicolet [Sat, 26 Sep 2020 21:19:53 +0000 (23:19 +0200)]
Tweak nnue scaling to keep more material

Current master uses a constant scale factor of 5/4 = 1.25 for the output
of the NNUE network, for compatibility with search and classical evaluation.
We modify this scale factor to make it dependent on the phase of the game,
going from about 1.5 in the opening to 1.0 for pure pawn endgames.

This helps Stockfish to avoid exchanges of pieces (heavy pieces in particular)
when she has the advantage, keeping more material on the board when attacking.

Passed STC:
LLR: 2.95 (-2.94,2.94) {-0.25,1.25}
Total: 14744 W: 1771 L: 1599 D: 11374
Ptnml(0-2): 87, 1184, 4664, 1344, 93
https://tests.stockfishchess.org/tests/view/5f6fb0a63b22d6afa506904f

Passed LTC:
LLR: 2.95 (-2.94,2.94) {0.25,1.25}
Total: 8912 W: 512 L: 393 D: 8007
Ptnml(0-2): 7, 344, 3637, 459, 9
https://tests.stockfishchess.org/tests/view/5f6fcf533b22d6afa5069066

closes https://github.com/official-stockfish/Stockfish/pull/3154

Bench: 3943952

14 months agoSwitch to NNUE eval probabilistically for OCB
SFisGOD [Wed, 23 Sep 2020 06:00:42 +0000 (14:00 +0800)]
Switch to NNUE eval probabilistically for OCB

Introduce a small chance of switching to NNUE if PSQ imbalance is large but we have opposite colored bishops and the classical eval is struggling to win.

STC:
LLR: 2.94 (-2.94,2.94) {-0.25,1.25}
Total: 25304 W: 3179 L: 2983 D: 19142
Ptnml(0-2): 172, 2171, 7781, 2345, 183
https://tests.stockfishchess.org/tests/view/5f6b14dec7759d4ee307cfe3

LTC:
LLR: 2.94 (-2.94,2.94) {0.25,1.25}
Total: 84680 W: 4846 L: 4556 D: 75278
Ptnml(0-2): 89, 3933, 34011, 4213, 94
https://tests.stockfishchess.org/tests/view/5f6b3fb6c7759d4ee307cff9

closes https://github.com/official-stockfish/Stockfish/pull/3146

Bench: 3865413

14 months agoSuppress info strings before 'uci'
Stéphane Nicolet [Thu, 24 Sep 2020 09:38:35 +0000 (11:38 +0200)]
Suppress info strings before 'uci'

On Windows, Stockfish wouldn't launch in some GUI because we output some
info strings (about the use of large pages) before sending the 'uci'
command. It seems more robust to suppress these info strings, and instead
to add a proper section section in the Readme about large pages use.

fixes https://github.com/official-stockfish/Stockfish/issues/3052
closes https://github.com/official-stockfish/Stockfish/pull/3147

No functional change

14 months agoIncrease reductions with the number of threads
Stéphane Nicolet [Tue, 22 Sep 2020 20:43:41 +0000 (22:43 +0200)]
Increase reductions with the number of threads

Passed STC with 8 threads:
LLR: 2.92 (-2.94,2.94) {-0.25,1.25}
Total: 13520 W: 1135 L: 1012 D: 11373
Ptnml(0-2): 39, 815, 4929, 938, 39
https://tests.stockfishchess.org/tests/view/5f68e274ded68c240be73f41

Passed LTC with 8 threads:
LLR: 2.96 (-2.94,2.94) {0.25,1.25}
Total: 48384 W: 2183 L: 1994 D: 44207
Ptnml(0-2): 28, 1777, 20402, 1948, 37
https://tests.stockfishchess.org/tests/view/5f68f068ded68c240be747e9

closes https://github.com/official-stockfish/Stockfish/pull/3142

No functional change (for one thread)

14 months agoSmall cleanups 12
Stéphane Nicolet [Wed, 9 Sep 2020 08:49:31 +0000 (10:49 +0200)]
Small cleanups 12

- Clean signature of functions in namespace NNUE
- Add comment for countermove based pruning
- Remove bestMoveCount variable
- Add const qualifier to kpp_board_index array
- Fix spaces in get_best_thread()
- Fix indention in capture LMR code in search.cpp
- Rename TtmemDeleter to LargePageDeleter

Closes https://github.com/official-stockfish/Stockfish/pull/3063

No functional change

14 months agoAdd large page support for NNUE weights and simplify TT mem management
Sami Kiminki [Sun, 30 Aug 2020 16:41:30 +0000 (19:41 +0300)]
Add large page support for NNUE weights and simplify TT mem management

Use TT memory functions to allocate memory for the NNUE weights. This
should provide a small speed-up on systems where large pages are not
automatically used, including Windows and some Linux distributions.

Further, since we now have a wrapper for std::aligned_alloc(), we can
simplify the TT memory management a bit:

- We no longer need to store separate pointers to the hash table and
  its underlying memory allocation.
- We also get to merge the Linux-specific and default implementations
  of aligned_ttmem_alloc().

Finally, we'll enable the VirtualAlloc code path with large page
support also for Win32.

STC: https://tests.stockfishchess.org/tests/view/5f66595823a84a47b9036fba
LLR: 2.94 (-2.94,2.94) {-0.25,1.25}
Total: 14896 W: 1854 L: 1686 D: 11356
Ptnml(0-2): 65, 1224, 4742, 1312, 105

closes https://github.com/official-stockfish/Stockfish/pull/3081

No functional change.

14 months agoTweak hybrid treshold.
Stefan Geschwentner [Sun, 20 Sep 2020 20:25:19 +0000 (22:25 +0200)]
Tweak hybrid treshold.

Increase the first hybrid threshold with more material.
Rewrite the hybrid rules for clarity.

STC:
LLR: 2.94 (-2.94,2.94) {-0.25,1.25}
Total: 24416 W: 3039 L: 2848 D: 18529
Ptnml(0-2): 135, 2136, 7503, 2271, 163
https://tests.stockfishchess.org/tests/view/5f6451efbb0cae038ca8f4dc

LTC;
LLR: 2.95 (-2.94,2.94) {0.25,1.25}
Total: 65016 W: 3702 L: 3455 D: 57859
Ptnml(0-2): 66, 2991, 26157, 3218, 76
https://tests.stockfishchess.org/tests/view/5f64b143bb0cae038ca8f51f

closes https://github.com/official-stockfish/Stockfish/pull/3140

Bench: 3973739

14 months agoSimplify reduced depth search
Unai Corzo [Sun, 20 Sep 2020 07:03:37 +0000 (09:03 +0200)]
Simplify reduced depth search

Simplification in reduced depth search.

STC https://tests.stockfishchess.org/tests/view/5f64c72fbb0cae038ca8f531
LLR: 2.94 (-2.94,2.94) {-1.25,0.25}
Total: 28320 W: 3475 L: 3359 D: 21486
Ptnml(0-2): 170, 2485, 8773, 2523, 209

LTC https://tests.stockfishchess.org/tests/view/5f650cfabb0cae038ca8f585
LLR: 2.95 (-2.94,2.94) {-0.75,0.25}
Total: 58392 W: 3354 L: 3285 D: 51753
Ptnml(0-2): 74, 2826, 23336, 2877, 83

closes https://github.com/official-stockfish/Stockfish/pull/3139

bench: 4201295

14 months agoUse tiling to speed up accumulator refreshes and updates
syzygy1 [Wed, 16 Sep 2020 15:39:11 +0000 (17:39 +0200)]
Use tiling to speed up accumulator refreshes and updates

Perform the update and refresh operations tile by tile in a local
array of vectors. By selecting the array size carefully, we
achieve that the compiler keeps the whole array in vector registers.

Idea and original implementation by @sf-x.

STC: https://tests.stockfishchess.org/tests/view/5f623eec912c15f19854b855
LLR: 2.94 (-2.94,2.94) {-0.25,1.25}
Total: 4872 W: 623 L: 477 D: 3772
Ptnml(0-2): 14, 350, 1585, 450, 37

LTC: https://tests.stockfishchess.org/tests/view/5f62434e912c15f19854b860
LLR: 2.94 (-2.94,2.94) {0.25,1.25}
Total: 25808 W: 1565 L: 1401 D: 22842
Ptnml(0-2): 23, 1186, 10332, 1330, 33

closes https://github.com/official-stockfish/Stockfish/pull/3130

No functional change

14 months agoSimplify futility pruning for captures
Unai Corzo [Wed, 16 Sep 2020 18:42:38 +0000 (20:42 +0200)]
Simplify futility pruning for captures

STC https://tests.stockfishchess.org/tests/view/5f61f0e4b91f2ec371e429c2
LLR: 2.94 (-2.94,2.94) {-1.25,0.25}
Total: 75512 W: 8747 L: 8704 D: 58061
Ptnml(0-2): 440, 6589, 23683, 6576, 468

LTC https://tests.stockfishchess.org/tests/view/5f6215d3912c15f19854b801
LLR: 2.95 (-2.94,2.94) {-0.75,0.25}
Total: 92912 W: 5030 L: 4992 D: 82890
Ptnml(0-2): 88, 4363, 37532, 4369, 104

closes https://github.com/official-stockfish/Stockfish/pull/3129

bench: 3856086

14 months agoRemove castling extension
Unai Corzo [Wed, 16 Sep 2020 17:14:32 +0000 (19:14 +0200)]
Remove castling extension

STC https://tests.stockfishchess.org/tests/view/5f5fa5348fbc1c8a3f476eca
LLR: 2.94 (-2.94,2.94) {-1.25,0.25}
Total: 38520 W: 4713 L: 4610 D: 29197
Ptnml(0-2): 233, 3486, 11734, 3559, 248

LTC https://tests.stockfishchess.org/tests/view/5f62166a912c15f19854b806
LLR: 2.93 (-2.94,2.94) {-0.75,0.25}
Total: 48024 W: 2673 L: 2600 D: 42751
Ptnml(0-2): 64, 2247, 19316, 2322, 63

closes https://github.com/official-stockfish/Stockfish/pull/3128

bench: 3818400

14 months agoAdded FEN string to bench output
GoldenRare [Thu, 10 Sep 2020 04:24:40 +0000 (00:24 -0400)]
Added FEN string to bench output

fixes https://github.com/official-stockfish/Stockfish/pull/3117

closes https://github.com/official-stockfish/Stockfish/pull/3118

No functional change

14 months agoImprove NDK section in Makefile
syzygy1 [Sun, 13 Sep 2020 18:16:52 +0000 (20:16 +0200)]
Improve NDK section in Makefile

This PR sets the "comp" variable simply to "clang",
which seems to be more consistent and allows a small simplification.

The PR also moves the section that sets "profile_make" and "profile_use" to after the NDK section,
which ensures that these variables are now set correctly for NDK/clang.

closes https://github.com/official-stockfish/Stockfish/pull/3121

No functional change

14 months agoUse 2 * bestMoveChanges.
xoto10 [Thu, 10 Sep 2020 20:10:57 +0000 (21:10 +0100)]
Use 2 * bestMoveChanges.

NNUE appears to provide a more stable eval than the classic eval,
so the time use dependencies on bestMoveChanges, fallingEval,
etc may need to change to make the best use of available time.
This change doubles the effect of totBestMoveChanges when giving
more time because the choice of best move is unstable.

STC:
LLR: 2.94 (-2.94,2.94) {-0.25,1.25}
Total: 101928 W: 11995 L: 11698 D: 78235 Elo +0.78
Ptnml(0-2): 592, 8707, 32103, 8936, 626
https://tests.stockfishchess.org/tests/view/5f538a462d02727c56b36cec

LTC:
LLR: 2.94 (-2.94,2.94) {0.25,1.25}
Total: 186392 W: 10383 L: 9877 D: 166132 Elo +0.81
Ptnml(0-2): 207, 8370, 75539, 8870, 210
https://tests.stockfishchess.org/tests/view/5f54a9712d02727c56b36d5a

closes https://github.com/official-stockfish/Stockfish/pull/3119

Bench 4222126

14 months agoUpdate default net to nn-03744f8d56d8.nnue
Sergio Vieri [Mon, 14 Sep 2020 09:24:05 +0000 (17:24 +0800)]
Update default net to nn-03744f8d56d8.nnue

Equivalent to 20200914-1520

closes https://github.com/official-stockfish/Stockfish/pull/3123

Bench: 4222126

14 months agoSimplify StatSCore Initialization
mckx00 [Mon, 14 Sep 2020 02:28:32 +0000 (19:28 -0700)]
Simplify StatSCore Initialization

No need to initialize StatScore at rootNode. Current Logic is redundant because at subsequent levels the grandchildren statScore is initialized to zero.

closes https://github.com/official-stockfish/Stockfish/pull/3122

Non functional change.

14 months agoDouble probability of using classical eval
SFisGOD [Sun, 6 Sep 2020 20:54:26 +0000 (04:54 +0800)]
Double probability of using classical eval

This patch doubles the moderate imbalance threshold and probability of using classical eval.
So now if imbalance is greater than PawnValueMg / 4 then there is a 1/8 chance of using classical eval.

STC:
LLR: 2.93 (-2.94,2.94) {-0.25,1.25}
Total: 10984 W: 1303 L: 1140 D: 8541
Ptnml(0-2): 58, 867, 3489, 1010, 68
https://tests.stockfishchess.org/tests/view/5f554c9f97da2d5437d3813e

LTC:
LLR: 2.95 (-2.94,2.94) {0.25,1.25}
Total: 43064 W: 2476 L: 2276 D: 38312
Ptnml(0-2): 37, 1985, 17308, 2145, 57
https://tests.stockfishchess.org/tests/view/5f55690a00a0aa2ca79f0a43

closes https://github.com/official-stockfish/Stockfish/pull/3114

Bench: 4161067

14 months agoAlways re-enable NNUE after "bench".
Gian-Carlo Pascutto [Tue, 8 Sep 2020 13:37:53 +0000 (15:37 +0200)]
Always re-enable NNUE after "bench".

Restore the default NNUE setting (enabled) after a bench command.
This also makes the resulting program settings independent of the
number of FENs that are being benched.

Fixes issue #3112.

closes https://github.com/official-stockfish/Stockfish/pull/3113

No functional change.

14 months agoBug fix in do_null_move() and NNUE simplification.
syzygy1 [Sun, 6 Sep 2020 15:29:12 +0000 (17:29 +0200)]
Bug fix in do_null_move() and NNUE simplification.

This fixes #3108 and removes some NNUE code that is currently not used.

At the moment, do_null_move() copies the accumulator from the previous
state into the new state, which is correct. It then clears the "computed_score"
flag because the side to move has changed, and with the other side to move
NNUE will return a completely different evaluation (normally with changed
sign but also with different NNUE-internal tempo bonus).

The problem is that do_null_move() clears the wrong flag. It clears the
computed_score flag of the old state, not of the new state. It turns out
that this almost never affects the search. For example, fixing it does not
change the current bench (but it does change the previous bench). This is
because the search code usually avoids calling evaluate() after a null move.

This PR corrects do_null_move() by removing the computed_score flag altogether.
The flag is not needed because nnue_evaluate() is never called twice on a position.

This PR also removes some unnecessary {}s and inserts a few blank lines
in the modified NNUE files in line with SF coding style.

Resulf ot STC non-regression test:
LLR: 2.95 (-2.94,2.94) {-1.25,0.25}
Total: 26328 W: 3118 L: 3012 D: 20198
Ptnml(0-2): 126, 2208, 8397, 2300, 133
https://tests.stockfishchess.org/tests/view/5f553ccc2d02727c56b36db1

closes https://github.com/official-stockfish/Stockfish/pull/3109

bench: 4109324

14 months agoUse classical eval more often
SFisGOD [Fri, 4 Sep 2020 09:14:50 +0000 (17:14 +0800)]
Use classical eval more often

If there is a moderate imbalance, use classical eval with small probability (1/16),
as derived from the node counter.

STC:
LLR: 2.94 (-2.94,2.94) {-0.25,1.25}
Total: 32320 W: 3562 L: 3377 D: 25381
Ptnml(0-2): 144, 2609, 10478, 2776, 153
https://tests.stockfishchess.org/tests/view/5f520615ba100690c5cc5f80

LTC:
LLR: 2.95 (-2.94,2.94) {0.25,1.25}
Total: 21032 W: 1116 L: 974 D: 18942
Ptnml(0-2): 20, 837, 8664, 971, 24
https://tests.stockfishchess.org/tests/view/5f522eaaba100690c5cc5f8c

closes https://github.com/official-stockfish/Stockfish/pull/3107

Bench: 4109324

14 months agoAdjust penalty on refuted early quiet moves
Vizvezdenec [Fri, 4 Sep 2020 12:53:59 +0000 (15:53 +0300)]
Adjust penalty on refuted early quiet moves

This patch changes how previous early moves are penalized in case
search finds a best move. Here, the first quiet move that was not
a transposition table move is penalized.

passed STC
https://tests.stockfishchess.org/tests/view/5f51d839ba100690c5cc5f69
LLR: 2.94 (-2.94,2.94) {-0.25,1.25}
Total: 10088 W: 1150 L: 997 D: 7941
Ptnml(0-2): 41, 772, 3278, 899, 54

passed LTC
https://tests.stockfishchess.org/tests/view/5f51e435ba100690c5cc5f76
LLR: 2.93 (-2.94,2.94) {0.25,1.25}
Total: 30808 W: 1564 L: 1405 D: 27839
Ptnml(0-2): 19, 1245, 12717, 1404, 19

closes https://github.com/official-stockfish/Stockfish/pull/3106

bench 3983758

14 months agoUpdate default net to nn-308d71810dff.nnue
Sergio Vieri [Thu, 3 Sep 2020 12:22:51 +0000 (20:22 +0800)]
Update default net to nn-308d71810dff.nnue

equivalent to 20200903-1739

Net trained from scratch, so it has quite different features extracted compared to the previous net (82215d0fd0df).

STC:
LLR: 2.98 (-2.94,2.94) {-0.25,1.25}
Total: 108328 W: 14048 L: 13719 D: 80561
Ptnml(0-2): 842, 10039, 32062, 10390, 831
https://tests.stockfishchess.org/tests/view/5f50e053ba100690c5cc5f00

LTC:
LLR: 2.96 (-2.94,2.94) {0.25,1.25}
Total: 13872 W: 1059 L: 890 D: 11923
Ptnml(0-2): 30, 724, 5270, 871, 41
https://tests.stockfishchess.org/tests/view/5f51821fba100690c5cc5f36

closes https://github.com/official-stockfish/Stockfish/pull/3104

Bench: 3832716

14 months agoLMR Simplification
VoyagerOne [Wed, 2 Sep 2020 20:58:44 +0000 (16:58 -0400)]
LMR Simplification

remove reduction at non-check cut nodes for second move at low depths

STC:
LLR: 2.95 (-2.94,2.94) {-1.25,0.25}
Total: 61712 W: 6594 L: 6543 D: 48575
Ptnml(0-2): 293, 5085, 20082, 5070, 326
https://tests.stockfishchess.org/tests/view/5f5007d6ba100690c5cc5ea9

LTC:
LLR: 2.94 (-2.94,2.94) {-0.75,0.25}
Total: 57544 W: 2983 L: 2925 D: 51636
Ptnml(0-2): 47, 2568, 23495, 2604, 58
https://tests.stockfishchess.org/tests/view/5f50c597ba100690c5cc5ef7

closes https://github.com/official-stockfish/Stockfish/pull/3103

Bench: 3952302

14 months agoSimplify singularQuietLMR
Unai Corzo [Thu, 3 Sep 2020 10:18:42 +0000 (12:18 +0200)]
Simplify singularQuietLMR

remove formerPV dependence

STC https://tests.stockfishchess.org/tests/view/5f4cb922ba100690c5cc5d35
LLR: 2.96 (-2.94,2.94) {-1.25,0.25}
Total: 113672 W: 12347 L: 12368 D: 88957
Ptnml(0-2): 566, 9537, 36699, 9420, 614

LTC https://tests.stockfishchess.org/tests/view/5f4e8474ba100690c5cc5e12
LLR: 2.93 (-2.94,2.94) {-0.75,0.25}
Total: 43032 W: 2298 L: 2227 D: 38507
Ptnml(0-2): 45, 1940, 17475, 2011, 45

closes https://github.com/official-stockfish/Stockfish/pull/3102

bench: 3290084

14 months agoLess pruning in qsearch
Unai Corzo [Wed, 2 Sep 2020 14:45:49 +0000 (16:45 +0200)]
Less pruning in qsearch

do not prune moves that give discovery checks, even if with negative SSE.

STC https://tests.stockfishchess.org/tests/view/5f4cb5e8ba100690c5cc5d25
LLR: 2.96 (-2.94,2.94) {-0.25,1.25}
Total: 91328 W: 9940 L: 9667 D: 71721
Ptnml(0-2): 491, 7345, 29693, 7670, 465

LTC https://tests.stockfishchess.org/tests/view/5f4dbc2eba100690c5cc5dac
LLR: 2.97 (-2.94,2.94) {0.25,1.25}
Total: 52448 W: 2799 L: 2586 D: 47063
Ptnml(0-2): 53, 2220, 21459, 2445, 47

closes https://github.com/official-stockfish/Stockfish/pull/3098

bench: 4031192

14 months agoRestore development version
Joost VandeVondele [Fri, 4 Sep 2020 05:46:06 +0000 (07:46 +0200)]
Restore development version

have fun!

No functional change

15 months agoStockfish 12
Joost VandeVondele [Wed, 2 Sep 2020 07:12:04 +0000 (09:12 +0200)]
Stockfish 12

Official release version of Stockfish 12

Bench: 3624569

-----------------------

It is our pleasure to release Stockfish 12 to users world-wide

Downloads will be freely available at

https://stockfishchess.org/download/

This version 12 of Stockfish plays significantly stronger than
any of its predecessors. In a match against Stockfish 11,
Stockfish 12 will typically win at least ten times more game pairs
than it loses.

This jump in strength, visible in regular progression tests during
development[1], results from the introduction of an efficiently
updatable neural network (NNUE) for the evaluation in Stockfish[2],
and associated tuning of the engine as a whole. The concept of the
NNUE evaluation was first introduced in shogi, and ported to
Stockfish afterward. Stockfish remains a CPU-only engine, since the
NNUE networks can be very efficiently evaluated on CPUs. The
recommended parameters of the NNUE network are embedded in
distributed binaries, and Stockfish will use NNUE by default.

Both the NNUE and the classical evaluations are available, and
can be used to assign values to positions that are later used in
alpha-beta (PVS) search to find the best move. The classical
evaluation computes this value as a function of various chess
concepts, handcrafted by experts, tested and tuned using fishtest.
The NNUE evaluation computes this value with a neural network based
on basic inputs. The network is optimized and trained on the
evaluations of millions of positions.

The Stockfish project builds on a thriving community of enthusiasts
that contribute their expertise, time, and resources to build a free
and open source chess engine that is robust, widely available, and
very strong. We invite chess fans to join the fishtest testing
framework and programmers to contribute on github[3].

Stay safe and enjoy chess!

The Stockfish team

[1] https://github.com/glinscott/fishtest/wiki/Regression-Tests
[2] https://github.com/official-stockfish/Stockfish/commit/84f3e867903f62480c33243dd0ecbffd342796fc
[3] https://stockfishchess.org/get-involved/

15 months agoUpdate CPU contributors list
Joost VandeVondele [Wed, 2 Sep 2020 06:05:08 +0000 (08:05 +0200)]
Update CPU contributors list

with fishtest data of Sept. 2 2020

closes https://github.com/official-stockfish/Stockfish/pull/3095

No functional change

15 months agoOnly use MADV_RANDOM if defined
Joost VandeVondele [Tue, 1 Sep 2020 10:22:47 +0000 (12:22 +0200)]
Only use MADV_RANDOM if defined

needed to compile on Haiku.

fixes https://github.com/official-stockfish/Stockfish/issues/3093

closes https://github.com/official-stockfish/Stockfish/pull/3094

No functional change

15 months agoLMR Root Node Simplification
VoyagerOne [Sun, 30 Aug 2020 17:57:57 +0000 (13:57 -0400)]
LMR Root Node Simplification

Simplify LMR at Root node

STC:
LLR: 2.94 (-2.94,2.94) {-1.25,0.25}
Total: 71520 W: 7649 L: 7614 D: 56257
Ptnml(0-2): 346, 5845, 23349, 5868, 352
https://tests.stockfishchess.org/tests/view/5f4be8c0ba100690c5cc5cbb

LTC:
LLR: 2.95 (-2.94,2.94) {-0.75,0.25}
Total: 74832 W: 3997 L: 3948 D: 66887
Ptnml(0-2): 77, 3422, 30362, 3485, 70
https://tests.stockfishchess.org/tests/view/5f4c603eba100690c5cc5d0e

closes https://github.com/official-stockfish/Stockfish/pull/3091

Bench: 3624569

15 months agoAlways print an info line before a bestmove
Joost VandeVondele [Mon, 31 Aug 2020 20:53:20 +0000 (22:53 +0200)]
Always print an info line before a bestmove

if very few nodes are being searched before a bestmove is reported,
an info line might be missing.

fixes https://github.com/official-stockfish/Stockfish/issues/2757

closes https://github.com/official-stockfish/Stockfish/pull/3088

no functional change

15 months agoUse llvm linker with clang on windows for LTO.
mstembera [Mon, 31 Aug 2020 03:48:10 +0000 (20:48 -0700)]
Use llvm linker with clang on windows for LTO.

other linkers might fail to link during the LTO phase.

The linker might have to be installed using
`pacman -Syu mingw-w64-x86_64-lld`

closes https://github.com/official-stockfish/Stockfish/pull/3084

No functional change.

15 months agoUse stable sort to make sure bench with TB yields same results everywhere.
mstembera [Mon, 31 Aug 2020 01:40:49 +0000 (18:40 -0700)]
Use stable sort to make sure bench with TB yields same results everywhere.

std::sort() is not stable so different implementations can produce different results:
use the stable version instead.

Observed for '8/6k1/5r2/8/8/8/1K6/Q7 w - - 0 1' yielding different bench results for gcc and MSVC
and 3-4-5 syzygy TB prior to this patch.

closes https://github.com/official-stockfish/Stockfish/pull/3083

No functional change.

15 months agoUpdate parametes in classical evaluation.
FauziAkram [Sun, 30 Aug 2020 10:58:05 +0000 (13:58 +0300)]
Update parametes in classical evaluation.

Passed STC (NNUE=False):
https://tests.stockfishchess.org/tests/view/5f42edfe5089a564a10d84a0
LLR: 2.96 (-2.94,2.94) {-0.25,1.25}
Total: 13840 W: 2591 L: 2336 D: 8913
Ptnml(0-2): 194, 1453, 3387, 1676, 210

Passed LTC (NNUE=False):
https://tests.stockfishchess.org/tests/view/5f4369795089a564a10d84d8
LLR: 2.95 (-2.94,2.94) {0.25,1.25}
Total: 159744 W: 19430 L: 18850 D: 121464
Ptnml(0-2): 960, 14185, 49030, 14709, 988

closes https://github.com/official-stockfish/Stockfish/pull/3080

bench: 3736029

15 months agoQS Pruning Simplification
VoyagerOne [Sun, 30 Aug 2020 01:13:05 +0000 (21:13 -0400)]
QS Pruning Simplification

Remove depth dependence in QS pruning

STC:
LLR: 2.95 (-2.94,2.94) {-1.25,0.25}
Total: 40536 W: 4442 L: 4358 D: 31736
Ptnml(0-2): 209, 3330, 13118, 3390, 221
https://tests.stockfishchess.org/tests/view/5f49035b3def6407861152f9

LTC:
LLR: 2.95 (-2.94,2.94) {-0.75,0.25}
Total: 97104 W: 5164 L: 5130 D: 86810
Ptnml(0-2): 103, 4478, 39377, 4470, 124
https://tests.stockfishchess.org/tests/view/5f4939d53def640786115322

closes https://github.com/official-stockfish/Stockfish/pull/3077

Bench: 3865238

15 months agoAdd / remove leaves from search tree ttPv
MJZ1977 [Fri, 28 Aug 2020 10:06:36 +0000 (12:06 +0200)]
Add / remove leaves from search tree ttPv

add if previous leaf is in search tree and we didn't find a counter move
else remove the position if the leaf is the last one in search tree.

STC : https://tests.stockfishchess.org/tests/view/5f49203c3def640786115314
LLR: 2.95 (-2.94,2.94) {-0.25,1.25}
Total: 29968 W: 3381 L: 3195 D: 23392
Ptnml(0-2): 146, 2432, 9671, 2560, 175

LTC : https://tests.stockfishchess.org/tests/view/5f494bea3def640786115336
LLR: 2.96 (-2.94,2.94) {0.25,1.25}
Total: 84952 W: 4619 L: 4333 D: 76000
Ptnml(0-2): 86, 3765, 34481, 4065, 79

closes https://github.com/official-stockfish/Stockfish/pull/3075

Bench 3527337

15 months agoReintroduce depth reduction
Unai Corzo [Fri, 28 Aug 2020 07:27:15 +0000 (09:27 +0200)]
Reintroduce depth reduction

Reintroduce depth reduction if the position is not in TT.

STC https://tests.stockfishchess.org/tests/view/5f4652e85089a564a10d868c
LLR: 2.97 (-2.94,2.94) {-0.25,1.25}
Total: 40240 W: 4535 L: 4331 D: 31374
Ptnml(0-2): 215, 3276, 12969, 3410, 250

LTC https://tests.stockfishchess.org/tests/view/5f46ca5e5089a564a10d86f3
LLR: 2.93 (-2.94,2.94) {0.25,1.25}
Total: 63096 W: 3426 L: 3188 D: 56482
Ptnml(0-2): 51, 2798, 25645, 2970, 84

closes https://github.com/official-stockfish/Stockfish/pull/3072

bench: 3611906

15 months agoCleaner make help
Joost VandeVondele [Wed, 26 Aug 2020 16:00:54 +0000 (18:00 +0200)]
Cleaner make help

do not print details if ARCH is an empty string. Follow up for b0b4ca17db49ed03057b5fa4ee4a12dab0e9c9e6

https://github.com/official-stockfish/Stockfish/pull/3071

No functional change

15 months agoEmbed default net, and simplify using non-default nets
Stéphane Nicolet [Sun, 23 Aug 2020 11:43:38 +0000 (13:43 +0200)]
Embed default net, and simplify using non-default nets

covers the most important cases from the user perspective:

It embeds the default net in the binary, so a download of that binary will result
in a working engine with the default net. The engine will be functional in the default mode
without any additional user action.

It allows non-default nets to be used, which will be looked for in up to
three directories (working directory, location of the binary, and optionally a specific default directory).
This mechanism is also kept for those developers that use MSVC,
the one compiler that doesn't have an easy mechanism for embedding data.

It is possible to disable embedding, and instead specify a specific directory, e.g. linux distros might want to use
CXXFLAGS="-DNNUE_EMBEDDING_OFF -DDEFAULT_NNUE_DIRECTORY=/usr/share/games/stockfish/" make -j ARCH=x86-64 profile-build

passed STC non-regression:
https://tests.stockfishchess.org/tests/view/5f4a581c150f0aef5f8ae03a
LLR: 2.95 (-2.94,2.94) {-1.25,-0.25}
Total: 66928 W: 7202 L: 7147 D: 52579
Ptnml(0-2): 291, 5309, 22211, 5360, 293

closes https://github.com/official-stockfish/Stockfish/pull/3070

fixes https://github.com/official-stockfish/Stockfish/issues/3030

No functional change.

15 months agoSimplify MCP in QS
VoyagerOne [Tue, 25 Aug 2020 13:10:47 +0000 (09:10 -0400)]
Simplify MCP in QS

Simplify moveCount pruning in QS by removing depth dependency.

STC
LLR: 2.94 (-2.94,2.94) {-1.25,0.25}
Total: 42960 W: 4741 L: 4661 D: 33558
Ptnml(0-2): 218, 3574, 13804, 3678, 206
https://tests.stockfishchess.org/tests/view/5f42e3f75089a564a10d8493

LTC
LLR: 2.94 (-2.94,2.94) {-0.75,0.25}
Total: 66672 W: 3563 L: 3508 D: 59601
Ptnml(0-2): 71, 3064, 26996, 3149, 56
https://tests.stockfishchess.org/tests/view/5f4353285089a564a10d84d0

closes https://github.com/official-stockfish/Stockfish/pull/3067

Bench: 4074430

15 months agoRemove Reduce Depth
VoyagerOne [Sun, 23 Aug 2020 16:04:50 +0000 (12:04 -0400)]
Remove Reduce Depth

Remove Reduce Depth at PV nodes.

STC:
LLR: 2.94 (-2.94,2.94) {-1.25,0.25}
Total: 56760 W: 6299 L: 6236 D: 44225
Ptnml(0-2): 286, 4843, 18076, 4872, 303
https://tests.stockfishchess.org/tests/view/5f41356087a5c3c63d8f53c9

LTC:
LLR: 2.95 (-2.94,2.94) {-0.75,0.25}
Total: 17496 W: 954 L: 865 D: 15677
Ptnml(0-2): 13, 768, 7098, 855, 14
https://tests.stockfishchess.org/tests/view/5f41bb7687a5c3c63d8f53f9

closes https://github.com/official-stockfish/Stockfish/pull/3055

Bench: 3555051

15 months agoRemove EvalList
syzygy1 [Mon, 24 Aug 2020 00:29:38 +0000 (02:29 +0200)]
Remove EvalList

This patch removes the EvalList structure from the Position object and generally simplifies the interface between do_move() and the NNUE code.

The NNUE evaluation function first calculates the "accumulator". The accumulator consists of two halves: one for white's perspective, one for black's perspective.

If the "friendly king" has moved or the accumulator for the parent position is not available, the accumulator for this half has to be calculated from scratch. To do this, the NNUE node needs to know the positions and types of all non-king pieces and the position of the friendly king. This information can easily be obtained from the Position object.

If the "friendly king" has not moved, its half of the accumulator can be calculated by incrementally updating the accumulator for the previous position. For this, the NNUE code needs to know which pieces have been added to which squares and which pieces have been removed from which squares. In principle this information can be derived from the Position object and StateInfo struct (in the same way as undo_move() does this). However, it is probably a bit faster to prepare this information in do_move(), so I have kept the DirtyPiece struct. Since the DirtyPiece struct now stores the squares rather than "PieceSquare" indices, there are now at most three "dirty pieces" (previously two). A promotion move that captures a piece removes the capturing pawn and the captured piece from the board (to SQ_NONE) and moves the promoted piece to the promotion square (from SQ_NONE).

An STC test has confirmed a small speedup:

https://tests.stockfishchess.org/tests/view/5f43f06b5089a564a10d850a
LLR: 2.94 (-2.94,2.94) {-0.25,1.25}
Total: 87704 W: 9763 L: 9500 D: 68441
Ptnml(0-2): 426, 6950, 28845, 7197, 434

closes https://github.com/official-stockfish/Stockfish/pull/3068

No functional change

15 months agoCheck ARCH=.... variable
Joost VandeVondele [Mon, 24 Aug 2020 19:32:04 +0000 (21:32 +0200)]
Check ARCH=.... variable

to prevent user errors or generating untested code,
check explicitly that the ARCH variable is equivalent to a supported architecture
as listed in `make help`.

To nevertheless compile for an untested target the user can override the internal
variable, passing the undocumented `SUPPORTED_ARCH=true` to make.

closes https://github.com/official-stockfish/Stockfish/pull/3062

No functional change.

15 months agoAllow for VNNI256 compilation with g++-8
mstembera [Mon, 24 Aug 2020 10:38:01 +0000 (03:38 -0700)]
Allow for VNNI256 compilation with g++-8

explicitly pass needed -mavx512f -mavx512bw flags

closes https://github.com/official-stockfish/Stockfish/pull/3061

No functional change

15 months agoIntroduce countermove based pruning for qsearch
Vizvezdenec [Mon, 24 Aug 2020 05:04:16 +0000 (08:04 +0300)]
Introduce countermove based pruning for qsearch

This patch continues work of previous patch in introducing pruning heuristics in qsearch by analogy to main search, now with countermove based pruning.
Idea is that if move is late enough and is quite check (we do generate them in qsearch) and has bad enough countermove history - prune it.

passed STC
https://tests.stockfishchess.org/tests/view/5f41220287a5c3c63d8f53c5
LLR: 2.93 (-2.94,2.94) {-0.25,1.25}
Total: 35944 W: 4127 L: 3929 D: 27888
Ptnml(0-2): 196, 2970, 11459, 3134, 213

passed LTC
https://tests.stockfishchess.org/tests/view/5f41862f87a5c3c63d8f53e8
LLR: 2.95 (-2.94,2.94) {0.25,1.25}
Total: 138448 W: 7655 L: 7252 D: 123541
Ptnml(0-2): 145, 6247, 56043, 6638, 151

closes https://github.com/official-stockfish/Stockfish/pull/3058

Bench: 3610676

15 months agoAllow TT entries with key16==0 to be fetched
Sami Kiminki [Fri, 21 Aug 2020 09:12:39 +0000 (12:12 +0300)]
Allow TT entries with key16==0 to be fetched

Fix the issue where a TT entry with key16==0 would always be reported
as a miss. Instead, we'll use depth8 to detect whether the TT entry is
occupied. In order to do that, we'll change DEPTH_OFFSET to -7
(depth8==0) to distinguish between an unoccupied entry and the
otherwise lowest possible depth, i.e., DEPTH_NONE (depth8==1).

To prevent a performance regression, we'll reorder the TT entry fields
by the access order of TranspositionTable::probe(). Memory in general
works fastest when accessed in sequential order. We'll also match the
store order in TTEntry::save() with the entry field order, and
re-order the 'if-or' expressions in TTEntry::save() from the cheapest
to the most expensive.

Finally, as we now have a proper TT entry occupancy test, we'll fix a
minor corner case with hashfull reporting. To reproduce:
- Use a big hash
- Either:
  a. Start 31 very quick searches (this wraparounds generation to 0); or
  b. Force generation of the first search to 0.
- go depth infinite

Before the fix, hashfull would incorrectly report nearly full hash
immediately after the search start, since
TranspositionTable::hashfull() used to consider only the entry
generation and not whether the entry was actually occupied.

STC:
LLR: 2.95 (-2.94,2.94) {-0.25,1.25}
Total: 36848 W: 4091 L: 3898 D: 28859
Ptnml(0-2): 158, 2996, 11972, 3091, 207
https://tests.stockfishchess.org/tests/view/5f3f98d5dc02a01a0c2881f7

LTC:
LLR: 2.95 (-2.94,2.94) {0.25,1.25}
Total: 32280 W: 1828 L: 1653 D: 28799
Ptnml(0-2): 34, 1428, 13051, 1583, 44
https://tests.stockfishchess.org/tests/view/5f3fe77a87a5c3c63d8f5332

closes https://github.com/official-stockfish/Stockfish/pull/3048

Bench: 3760677

15 months agoSupport VNNI on 256bit vectors
mstembera [Thu, 20 Aug 2020 23:59:27 +0000 (16:59 -0700)]
Support VNNI on 256bit vectors

due to downclocking on current chips (tested up to cascade lake)
supporting avx512 and vnni512, it is better to use avx2 or vnni256
in multithreaded (in particular hyperthreaded) engine use.
In single threaded use, the picture is different.

gcc compilation for vnni256 requires a toolchain for gcc >= 9.

closes https://github.com/official-stockfish/Stockfish/pull/3038

No functional change

15 months agoarmv8 AArch64 does not require -mfpu=neon
George Sobala [Mon, 24 Aug 2020 05:37:42 +0000 (06:37 +0100)]
armv8 AArch64 does not require -mfpu=neon

-mpfu is not required on AArch64 / armv8 architecture on Linux and throws an error if present.
This PR has been tested on gcc and clang on Gentoo-64 and Raspian-64 on a Raspberry Pi 4,
as well as with a cross from Ubuntu
(`make clean && make -j build ARCH=armv8         COMP=gcc COMPILER=aarch64-linux-gnu-g++`)

fixes https://github.com/official-stockfish/Stockfish/issues/3056
closes https://github.com/official-stockfish/Stockfish/pull/3059

No functional change

15 months agoIntroduce movecount pruning for qsearch()
Vizvezdenec [Sun, 23 Aug 2020 11:22:32 +0000 (14:22 +0300)]
Introduce movecount pruning for qsearch()

If in quiescence search, we assume that me can prune late moves when:

a) the move ordering count of the move is : moveCount > abs(depth) + 2
b) we are not in check
c) the late move does not give check
d) the late move is not an advanced pawn push

Modification of an original idea by @VoyagerOne.

STC
https://tests.stockfishchess.org/tests/view/5f40581787a5c3c63d8f535f
LLR: 2.96 (-2.94,2.94) {-0.25,1.25}
Total: 132848 W: 14999 L: 14661 D: 103188
Ptnml(0-2): 684, 11242, 42309, 11430, 759

LTC
https://tests.stockfishchess.org/tests/view/5f4226da87a5c3c63d8f5412
LLR: 2.98 (-2.94,2.94) {0.25,1.25}
Total: 12008 W: 678 L: 551 D: 10779
Ptnml(0-2): 8, 485, 4899, 596, 16

closes https://github.com/official-stockfish/Stockfish/pull/3053

Bench: 3749974

15 months agoSkip the alignment bug workaround for Clang
syzygy1 [Sat, 22 Aug 2020 11:36:34 +0000 (13:36 +0200)]
Skip the alignment bug workaround for Clang

Clang-10.0.0 poses as gcc-4.2:

$ clang++ -E -dM - </dev/null | grep GNUC

This means that Clang is using the workaround for the alignment bug of gcc-8
even though it does not have the bug (as far as I know).

This patch should speed up AVX2 and AVX512 compiles on Windows (when using Clang),
because it disables (for Clang) the gcc workaround we had introduced in this commit:
https://github.com/official-stockfish/Stockfish/commit/875183b310a8249922c2155e82cb4cecfae2097e

closes https://github.com/official-stockfish/Stockfish/pull/3050

No functional change.

15 months agoSmall trivial cleanups
Joost VandeVondele [Sat, 11 Jul 2020 14:59:33 +0000 (16:59 +0200)]
Small trivial cleanups

closes https://github.com/official-stockfish/Stockfish/pull/2801

No functional change

15 months agoInstructions to build on older Macintosh
Stéphane Nicolet [Sat, 22 Aug 2020 09:37:53 +0000 (11:37 +0200)]
Instructions to build on older Macintosh

In recent Macs, it is possible to use the Clang compiler provided by Apple
to compile Stockfish out of the box, and this is the method used by default
in our Makefile (the Makefile sets the macosx-version-min=10.14 flag to select
the right libc++ library for the Clang compiler with recent c++17 support).

But it is quite possible to compile and run Stockfish on older Macs! Below
we describe a method to install a recent GNU compiler on these Macs, to get
the c++17 support. We have tested the following procedure to install gcc10 on
machines running Mac OS 10.7, Mac OS 10.9 and Mac OS 10.13:

1) install XCode for your machine.

2) install Apple command-line developer tools for XCode, by typing the following
   command in a Terminal:

```
      sudo xcode-select --install
```

3) go to the Stockfish "src" directory, then try a default build and run Stockfish:

```
      make clean
      make build
      make net
      ./stockfish
```

4) if step 3 worked, congrats! You have a compiler recent enough on your Mac
to compile Stockfish. If not, continue with step 5 to install GNU gcc10 :-)

5) install the MacPorts package manager (https://www.macports.org/install.php),
for instance using the fast method in the "macOS Package (.pkg) Installer"
section of the page.

6) use the "port" command to install the gcc10 package of MacPorts by typing the
following command:

```
    sudo port install gcc10
```

With this step, MacPorts will install the gcc10 compiler under the name "g++-mp-10"
in the /opt/local/bin directory:

```
   which g++-mp-10

   /opt/local/bin/g++-mp-10       <--- answer
```

7) You can now go back to the "src" directory of Stockfish, and try to build
Stockfish by pointing at the right compiler:

```
   make clean
   make build COMP=gcc COMPCXX=/opt/local/bin/g++-mp-10
   make net
   ./stockfish
```

8) Enjoy Stockfish on Macintosh!

See this pull request for further discussion:
https://github.com/official-stockfish/Stockfish/pull/3049

No functional change

15 months agoExplicitly rely on pthreads if possible
Joost VandeVondele [Fri, 21 Aug 2020 20:10:55 +0000 (22:10 +0200)]
Explicitly rely on pthreads if possible

allows us to set the needed stacksize on thread creation.

Useful for environments with too small a default stack size (e.g. Alpine Linux with musl).

Passed STC, no regression:

LLR: 2.96 (-2.94,2.94) {-1.25,0.25}
Total: 17816 W: 1344 L: 1275 D: 15197
Ptnml(0-2): 30, 1057, 6682, 1092, 47
https://tests.stockfishchess.org/tests/view/5f402b5587a5c3c63d8f534d

closes https://github.com/official-stockfish/Stockfish/pull/3047

fixes https://github.com/official-stockfish/Stockfish/issues/3041

No functional change.

15 months agoDisplay classic and NNUE evaluation in trace mode
MJZ1977 [Fri, 21 Aug 2020 08:57:34 +0000 (10:57 +0200)]
Display classic and NNUE evaluation in trace mode

show both the classical and NNUE evaluation,
as well as the Final evaluation.

closes https://github.com/official-stockfish/Stockfish/pull/3042

No functional change.

15 months agoSimplify away internal iterative deepening
Unai Corzo [Fri, 21 Aug 2020 07:24:25 +0000 (09:24 +0200)]
Simplify away internal iterative deepening

Remove the iterative deepening step.
Instead, employ a depth reduction if the position is not in TT and on the PV.

STC https://tests.stockfishchess.org/tests/view/5f3ce6eaa95672ddd56c637e
LLR: 2.97 (-2.94,2.94) {-0.50,1.50}
Total: 41096 W: 4421 L: 4257 D: 32418
Ptnml(0-2): 207, 3259, 13460, 3407, 215

LTC (old) https://tests.stockfishchess.org/tests/view/5f3d7d4fa95672ddd56c640b
LLR: 2.92 (-2.94,2.94) {-1.50,0.50}
Total: 26032 W: 1320 L: 1309 D: 23403
Ptnml(0-2): 22, 1152, 10654, 1169, 19

LTC (new) https://tests.stockfishchess.org/tests/view/5f3e31e0a95672ddd56c6464
LLR: 2.95 (-2.94,2.94) {-0.75,0.25}
Total: 34160 W: 1844 L: 1766 D: 30550
Ptnml(0-2): 33, 1533, 13876, 1599, 39

bench: 3849173

15 months agoUpdate Makefile for macOS
gsobala [Fri, 21 Aug 2020 10:28:53 +0000 (11:28 +0100)]
Update Makefile for macOS

Changes to deal with compilation (particularly profile-build) on macOS.
(1) The default toolchain has gcc masquerading as clang,
    the previous Makefile was not picking up the required changes
    to the different profiling tools.
(2) The previous Makefile test for gccisclang occurred before
    a potential overwrite of CXX by COMPCXX
(3) llvm-profdata no longer runs as a command on macOS and
    instead is invoked by ``xcrun llvm-profdata``
(4) Needs to support use of true gcc using e.g.
    ``make build ... COMPCXX=g++-10``
(5) enable profile-build in travis for macOS

closes https://github.com/official-stockfish/Stockfish/pull/3043

No functional change

15 months agoDeal with very old linux kernels
Joost VandeVondele [Fri, 21 Aug 2020 05:42:19 +0000 (07:42 +0200)]
Deal with very old linux kernels

MADV_HUGEPAGE might not be available, for kernels before 2.6.38 (released 2011). Just skip the madvise.

closes https://github.com/official-stockfish/Stockfish/pull/3039

No functional change

15 months agoSet Use NNUE by default to true
Joost VandeVondele [Thu, 20 Aug 2020 12:24:49 +0000 (14:24 +0200)]
Set Use NNUE by default to true

Since the initial stages of the merge, progress has been made so that
this seems the best option now:

* NNUE is clearly stronger on most relevant hardware and time controls
* All of our CI and testing infrastructure has been adjusted
* The default net is easy to get (further ideas #3030)

fixes https://github.com/official-stockfish/Stockfish/issues/2861

closes https://github.com/official-stockfish/Stockfish/pull/3033

No functional change.

15 months agoSend error message as an UCI info string
Joost VandeVondele [Wed, 19 Aug 2020 17:21:41 +0000 (19:21 +0200)]
Send error message as an UCI info string

some GUIs do not show the error message when the engine terminates in the no-net case, as it is send to cerr.
Instead send it as an info string, which the GUI will more likely display.

closes https://github.com/official-stockfish/Stockfish/pull/3031

No functional change.

15 months agoReintroduce last captures extension
SFisGOD [Tue, 18 Aug 2020 10:54:28 +0000 (18:54 +0800)]
Reintroduce last captures extension

STC:
LLR: 2.93 (-2.94,2.94) {-0.50,1.50}
Total: 34840 W: 3834 L: 3682 D: 27324
Ptnml(0-2): 153, 2767, 11455, 2865, 180
https://tests.stockfishchess.org/tests/view/5f3bb380b38d442594aabefc

LTC:
LLR: 2.95 (-2.94,2.94) {0.25,1.75}
Total: 15832 W: 890 L: 776 D: 14166
Ptnml(0-2): 17, 669, 6429, 785, 16
https://tests.stockfishchess.org/tests/view/5f3c46a0a95672ddd56c632a

closes https://github.com/official-stockfish/Stockfish/pull/3028

see also https://github.com/official-stockfish/Stockfish/pull/3020

Bench: 4348811

15 months agoExpanded support for x86-32 architectures.
syzygy1 [Mon, 17 Aug 2020 23:56:12 +0000 (01:56 +0200)]
Expanded support for x86-32 architectures.

add new ARCH targets

x86-32-sse41-popcnt     > x86 32-bit with sse41 and popcnt support
x86-32-sse2             > x86 32-bit with sse2 support
x86-32                  > x86 32-bit generic (with mmx and sse support)

retire x86-32-old (use general-32)

closes https://github.com/official-stockfish/Stockfish/pull/3022

No functional change.

15 months agoBetter error message on missing curl/wget
Joost VandeVondele [Tue, 18 Aug 2020 16:06:28 +0000 (18:06 +0200)]
Better error message on missing curl/wget

provide clean error/warning message for missing curl/wget, sha256sum/shasum

fixes https://github.com/official-stockfish/Stockfish/issues/3025

closes https://github.com/official-stockfish/Stockfish/pull/3026

No functional change

15 months agoFix Makefile typo
Joost VandeVondele [Tue, 18 Aug 2020 06:49:06 +0000 (08:49 +0200)]
Fix Makefile typo

remove stray quote, shown with `make help`

No functional change

15 months agoFallback to NNUE
mstembera [Sun, 16 Aug 2020 22:23:50 +0000 (15:23 -0700)]
Fallback to NNUE

If the classical eval ends up much smaller than estimated fall back to NNUE.
Also use multiply instead of divide for the threshold comparison for smoother transitions without rounding.

STC https://tests.stockfishchess.org/tests/view/5f3a5011b38d442594aabdfe
LLR: 2.96 (-2.94,2.94) {-0.50,1.50}
Total: 57352 W: 6325 L: 6135 D: 44892
Ptnml(0-2): 277, 4748, 18482, 4846, 323

LTC https://tests.stockfishchess.org/tests/view/5f3aee9db38d442594aabe82
LLR: 2.95 (-2.94,2.94) {0.25,1.75}
Total: 16232 W: 897 L: 781 D: 14554
Ptnml(0-2): 19, 679, 6616, 771, 31

closes https://github.com/official-stockfish/Stockfish/pull/3023

bench: 4026216

-----

Recommended net: https://tests.stockfishchess.org/api/nn/nn-82215d0fd0df.nnue

15 months agoRemove last captures extension
Unai Corzo [Mon, 17 Aug 2020 16:22:32 +0000 (18:22 +0200)]
Remove last captures extension

STC https://tests.stockfishchess.org/tests/view/5f395657e98b6c64b3df41dd
LLR: 2.95 (-2.94,2.94) {-1.50,0.50}
Total: 144664 W: 15426 L: 15537 D: 113701
Ptnml(0-2): 612, 11341, 48537, 11230, 612

LTC https://tests.stockfishchess.org/tests/view/5f3a2ec7b38d442594aabdd7
LLR: 2.96 (-2.94,2.94) {-1.50,0.50}
Total: 22728 W: 1161 L: 1146 D: 20421
Ptnml(0-2): 21, 960, 9388, 973, 22

closes https://github.com/official-stockfish/Stockfish/pull/3020

bench: 3832662

15 months agoRemove history bonus from Eval
VoyagerOne [Mon, 17 Aug 2020 12:58:03 +0000 (08:58 -0400)]
Remove history bonus from Eval

STC:
LLR: 2.92 (-2.94,2.94) {-1.50,0.50}
Total: 26776 W: 2787 L: 2725 D: 21264
https://tests.stockfishchess.org/tests/view/5f39d6beb38d442594aabd9b

LTC:
LLR: 2.93 (-2.94,2.94) {-1.50,0.50}
Total: 12968 W: 635 L: 608 D: 11725
https://tests.stockfishchess.org/tests/view/5f39decfb38d442594aabda7

closes https://github.com/official-stockfish/Stockfish/pull/3019

Bench:  4335100

15 months agoSupport building for Android using NDK
notruck [Sun, 16 Aug 2020 15:59:13 +0000 (08:59 -0700)]
Support building for Android using NDK

The easiest way to use the NDK in conjunction with this Makefile (tested on linux-x86_64):

1. Download the latest NDK (r21d) from Google from https://developer.android.com/ndk/downloads
2. Place and unzip the NDK in $HOME/ndk folder
3. Export the path variable e.g., `export PATH=$PATH:$HOME/ndk/android-ndk-r21d/toolchains/llvm/prebuilt/linux-x86_64/bin`
4. cd to your Stockfish/src dir
5. Issue `make -j ARCH=armv8 COMP=ndk build`  (use `ARCH=armv7` or `ARCH=armv7-neon` for older CPUs)
6. Optionally `make -j ARCH=armv8 COMP=ndk strip`
7. That's all. Enjoy!

Improves support from Raspberry Pi (incomplete?) and compiling on arm in general

closes https://github.com/official-stockfish/Stockfish/pull/3015

fixes https://github.com/official-stockfish/Stockfish/issues/2860

fixes https://github.com/official-stockfish/Stockfish/issues/2641

Support is still fragile as we're missing CI on these targets. Nevertheless tested with:

```bash
  # build crosses from ubuntu 20.04 on x86 to various arch/OS combos
  # tested with suitable packages installed
  # (build-essentials, mingw-w64, g++-arm-linux-gnueabihf, NDK (r21d) from google)

  # cross to Android
  export PATH=$HOME/ndk/android-ndk-r21d/toolchains/llvm/prebuilt/linux-x86_64/bin:$PATH
  make clean && make -j build ARCH=armv7         COMP=ndk  && make -j build ARCH=armv7 COMP=ndk strip
  make clean && make -j build ARCH=armv7-neon    COMP=ndk  && make -j build ARCH=armv7-neon COMP=ndk strip
  make clean && make -j build ARCH=armv8         COMP=ndk  && make -j build ARCH=armv8 COMP=ndk strip

  # cross to Raspberry Pi
  make clean && make -j build ARCH=armv7         COMP=gcc COMPILER=arm-linux-gnueabihf-g++
  make clean && make -j build ARCH=armv7-neon    COMP=gcc COMPILER=arm-linux-gnueabihf-g++

  # cross to Windows
  make clean && make -j build ARCH=x86-64-modern COMP=mingw
```

No functional change

15 months agoSimplify away the passed pawn extension
Unai Corzo [Mon, 17 Aug 2020 07:22:15 +0000 (09:22 +0200)]
Simplify away the passed pawn extension

STC https://tests.stockfishchess.org/tests/view/5f3955f0e98b6c64b3df41d7
LLR: 2.96 (-2.94,2.94) {-1.50,0.50}
Total: 31992 W: 3611 L: 3548 D: 24833
Ptnml(0-2): 174, 2658, 10273, 2713, 178

LTC https://tests.stockfishchess.org/tests/view/5f399e41e98b6c64b3df4210
LLR: 3.01 (-2.94,2.94) {-1.50,0.50}
Total: 29568 W: 1488 L: 1480 D: 26600
Ptnml(0-2): 40, 1272, 12142, 1300, 30

closes https://github.com/official-stockfish/Stockfish/pull/3017

bench: 3844671

-----

Recommended net: https://tests.stockfishchess.org/api/nn/nn-82215d0fd0df.nnue

15 months agoReformat code in little-endian patch
Stéphane Nicolet [Sun, 16 Aug 2020 19:46:54 +0000 (21:46 +0200)]
Reformat code in little-endian patch

Reformat code and rename the function to "read_little_endian()" in the recent
commit by Ronald de Man for support of big endian systems.

closes https://github.com/official-stockfish/Stockfish/pull/3016

No functional change
-----

Recommended net: https://tests.stockfishchess.org/api/nn/nn-82215d0fd0df.nnue

15 months agoAdd further targets to travis testing
Joost VandeVondele [Sun, 16 Aug 2020 11:21:07 +0000 (13:21 +0200)]
Add further targets to travis testing

general-32, general-64 and help

closes https://github.com/official-stockfish/Stockfish/pull/3014

No functional change

15 months agoAssume network file is in little-endian byte order
syzygy1 [Sat, 15 Aug 2020 14:50:39 +0000 (16:50 +0200)]
Assume network file is in little-endian byte order

This patch fixes the byte order when reading 16- and 32-bit values from the network file on a big-endian machine.

Bytes are ordered in read_le() using unsigned arithmetic, which doesn't need tricks to determine the endianness of the machine. Unfortunately the compiler doesn't seem to be able to optimise the ordering operation, but reading in the weights is not a time-critical operation and the extra time it takes should not be noticeable.

Big endian systems are still untested with NNUE.

fixes #3007

closes https://github.com/official-stockfish/Stockfish/pull/3009

No functional change.

15 months agoDisplay NEON in compiler string
Daylen Yang [Sat, 15 Aug 2020 02:53:46 +0000 (19:53 -0700)]
Display NEON in compiler string

if NEON intrinsics are being used and USE_NEON is defined.

closes https://github.com/official-stockfish/Stockfish/pull/3008

No functional change

15 months agoUse NNUE more for fortresses
Joost VandeVondele [Fri, 14 Aug 2020 20:18:12 +0000 (22:18 +0200)]
Use NNUE more for fortresses

Increases the use of NNUE evaluation in positions without captures/pawn moves,
by increasing the NNUEThreshold threshold with rule50_count.

This patch will force Stockfish to use NNUE eval more and more in materially
unbalanced positions, when it seems that the classical eval is struggling to
win and only manages to shuffle. This will ask the (slower) NNUE eval to
double-check the potential fortress branches of the search tree, but only
when necessary.

passed STC:
https://tests.stockfishchess.org/tests/view/5f36f1bf11a9b1a1dbf192d8
LLR: 2.93 (-2.94,2.94) {-0.50,1.50}
Total: 51824 W: 5836 L: 5653 D: 40335
Ptnml(0-2): 264, 4356, 16512, 4493, 287

passed LTC:
https://tests.stockfishchess.org/tests/view/5f37836111a9b1a1dbf1936d
LLR: 2.93 (-2.94,2.94) {0.25,1.75}
Total: 29768 W: 1747 L: 1590 D: 26431
Ptnml(0-2): 33, 1347, 11977, 1484, 43

closes https://github.com/official-stockfish/Stockfish/pull/3011

Bench: 4173967

15 months agoTry to match relative magnitude of NNUE eval to classical
mstembera [Fri, 14 Aug 2020 11:49:33 +0000 (04:49 -0700)]
Try to match relative magnitude of NNUE eval to classical

The idea is that since we are mixing NNUE and classical evals matching their magnitudes closer allows for better comparisons.

STC https://tests.stockfishchess.org/tests/view/5f35a65411a9b1a1dbf18e2b
LLR: 2.94 (-2.94,2.94) {-0.50,1.50}
Total: 9840 W: 1150 L: 1027 D: 7663
Ptnml(0-2): 49, 772, 3175, 855, 69

LTC https://tests.stockfishchess.org/tests/view/5f35bcbe11a9b1a1dbf18e47
LLR: 2.93 (-2.94,2.94) {0.25,1.75}
Total: 44424 W: 2492 L: 2294 D: 39638
Ptnml(0-2): 42, 2015, 17915, 2183, 57

also corrects the location to clamp the evaluation (non-function on bench).

closes https://github.com/official-stockfish/Stockfish/pull/3003

bench: 3905447

15 months agoAlso dampen NNUE eval with 50 move rule
Miguel Lahoz [Mon, 10 Aug 2020 14:57:11 +0000 (22:57 +0800)]
Also dampen NNUE eval with 50 move rule

Move the existing dampening function last so that NNUE evaluations are
also handled as we approach the 50 move rule.

STC:
LLR: 2.95 (-2.94,2.94) {-0.50,1.50}
Total: 4792 W: 695 L: 561 D: 3536
Ptnml(0-2): 19, 420, 1422, 478, 57
https://tests.stockfishchess.org/tests/view/5f3164179081672066537534

LTC:
LLR: 8.62 (-2.94,2.94) {0.25,1.75}
Total: 286744 W: 18494 L: 17430 D: 250820
Ptnml(0-2): 418, 14886, 111745, 15860, 463
https://tests.stockfishchess.org/tests/view/5f316b039081672066537541

closes https://github.com/official-stockfish/Stockfish/pull/3004

Bench: 4001800

15 months agoVerify SHA of downloaded net file
Joost VandeVondele [Thu, 13 Aug 2020 20:54:13 +0000 (22:54 +0200)]
Verify SHA of downloaded net file

check SHA of the available and downloaded file.

Document the format requirement on the default net.

Also allow curl to make possibly insecure connections, as needed for old curl.

fixes https://github.com/official-stockfish/Stockfish/issues/2998

closes https://github.com/official-stockfish/Stockfish/pull/3000

No functional change.

15 months agoClean-up Makefile help
Stéphane Nicolet [Thu, 13 Aug 2020 11:40:06 +0000 (13:40 +0200)]
Clean-up Makefile help

Do not show the details of the default architecture for a simple "make help"
invocation, as the details are most likely to confuse beginners. Instead we
make it clear which architecture is the default and put an example at the end
of the Makefile as an incentative to use "make help ARCH=blah" to discover
the flags used by the different architectures.

```
    make help
    make help ARCH=x86-64-ssse3
```

Also clean-up and modernize a bit the Makefile examples while at it.

closes https://github.com/official-stockfish/Stockfish/pull/2996

No functional change

15 months agoUpdate default net to nn-82215d0fd0df.nnue
Sergio Vieri [Wed, 12 Aug 2020 15:21:21 +0000 (23:21 +0800)]
Update default net to nn-82215d0fd0df.nnue

Net created at: 20200812-2257

passed STC: https://tests.stockfishchess.org/tests/view/5f340ca99e5f2effc089da17
LLR: 2.96 (-2.94,2.94) {-0.50,1.50}
Total: 5744 W: 756 L: 627 D: 4361
Ptnml(0-2): 28, 485, 1731, 586, 42

passed LTC: https://tests.stockfishchess.org/tests/view/5f341eba9e5f2effc089da23
LLR: 2.94 (-2.94,2.94) {0.25,1.75}
Total: 17136 W: 1041 L: 917 D: 15178
Ptnml(0-2): 13, 813, 6807, 907, 28

closes https://github.com/official-stockfish/Stockfish/pull/2992

Bench: 3935117

15 months agoOutput the SSE2 flag in compiler_info
Joost VandeVondele [Wed, 12 Aug 2020 15:21:12 +0000 (17:21 +0200)]
Output the SSE2 flag in compiler_info

was missing in the list of outputs, slightly reorder flags.
explicitly add -msse2 if USE_SSE2 (is implicit already, -msse -m64).

closes https://github.com/official-stockfish/Stockfish/pull/2990

No functional change.

15 months agoAdd support for VNNI
mstembera [Tue, 11 Aug 2020 19:59:39 +0000 (12:59 -0700)]
Add support for VNNI

Adds support for Vector Neural Network Instructions (avx512), as available on Intel Cascade Lake

The _mm512_dpbusd_epi32() intrinsic (vpdpbusd instruction) is taylor made for NNUE.

on a cascade lake CPU (AWS C5.24x.large, gcc 10) NNUE eval is at roughly 78% nps of classical
(single core test)

bench 1024 1 24 default depth:
target  classical  NNUE  ratio
vnni  2207232  1725987  78.20
avx512  2216789  1671734  75.41
avx2  2194006  1611263  73.44
modern  2185001  1352469  61.90

closes https://github.com/official-stockfish/Stockfish/pull/2987

No functional change

15 months agoUse posix_memalign for Apple Silicon instead of _mm_malloc
Daylen Yang [Tue, 11 Aug 2020 19:02:48 +0000 (12:02 -0700)]
Use posix_memalign for Apple Silicon instead of _mm_malloc

fails to build on that target, because of missing Intel Intrinsics.
macOS has posix_memalign() since ~2014 so we can simplify the code and just use that for all Apple platforms.

closes https://github.com/official-stockfish/Stockfish/pull/2985

No functional change.

15 months agoRestrict avx2 hack to windows target
Joost VandeVondele [Tue, 11 Aug 2020 19:11:17 +0000 (21:11 +0200)]
Restrict avx2 hack to windows target

this workaround is possibly rather a windows & gcc specific problem. See e.g.
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412#c25

on Linux with gcc 8 this patch brings roughly a 8% speedup.
However, probably needs some testing in the wild.

includes a workaround for an old msys make (3.81) installation (fixes #2984)

No functional change

15 months agoTweak castling extension
SFisGOD [Mon, 10 Aug 2020 13:13:56 +0000 (21:13 +0800)]
Tweak castling extension

Change condition from three friendly pieces to two. This now means that we only extend castling on the king side if there are no other friendly pieces aside from king and rook. For the queen side, we only extend if there is only a rook and another friendly piece or if there is only a single rook and no other friendly piece but this is very rare.

STC:
LLR: 3.20 (-2.94,2.94) {-0.50,1.50}
Total: 31144 W: 4086 L: 3903 D: 23155
Ptnml(0-2): 227, 2843, 9278, 2968, 256
https://tests.stockfishchess.org/tests/view/5f31487f9081672066537516

LTC:
LLR: 2.93 (-2.94,2.94) {0.25,1.75}
Total: 57816 W: 3786 L: 3538 D: 50492
Ptnml(0-2): 92, 2991, 22488, 3251, 86
https://tests.stockfishchess.org/tests/view/5f3167c3908167206653753d

closes https://github.com/official-stockfish/Stockfish/pull/2980

Bench: 4244812

15 months agoThis commit enables a mixed bench, to improve CI and allow for PGO (profile-build...
Guy Vreuls [Fri, 7 Aug 2020 15:07:46 +0000 (17:07 +0200)]
This commit enables a mixed bench, to improve CI and allow for PGO (profile-build) of the NNUE part of the code.

Joint work gvreuls / vondele

* Download the default NNUE net in AppVeyor
* Download net in travis CI `make net`
* Adjust tests to cover more archs, speedup instrumented testing
* Introduce 'mixed' bench as default, with further options:

classical, NNUE, mixed.

mixed (default) and NNUE require the default net to be present,
which can be obtained with

```
make net
```

Further examples (first is equivalent to `./stockfish bench`):

```
./stockfish bench 16 1 13 default depth mixed
./stockfish bench 16 1 13 default depth classical
./stockfish bench 16 1 13 default depth NNUE
```

The net is now downloaded automatically if needed for `profile-build`
(usual `build` works fine without net present)

PGO gives a nice speedup on fishtest:

passed STC:
LLR: 2.93 (-2.94,2.94) {-0.50,1.50}
Total: 3360 W: 469 L: 343 D: 2548
Ptnml(0-2): 20, 246, 1030, 356, 28
https://tests.stockfishchess.org/tests/view/5f31b5499081672066537569

passed LTC:
LLR: 2.97 (-2.94,2.94) {0.25,1.75}
Total: 8824 W: 609 L: 502 D: 7713
Ptnml(0-2): 8, 430, 3438, 519, 17
https://tests.stockfishchess.org/tests/view/5f31c87b908167206653757c

closes https://github.com/official-stockfish/Stockfish/pull/2931

fixes https://github.com/official-stockfish/Stockfish/issues/2907

requires fishtest updates before commit

Bench: 4290577

15 months agoFix AVX512 build with older compilers
mstembera [Mon, 10 Aug 2020 19:52:46 +0000 (12:52 -0700)]
Fix AVX512 build with older compilers

avoids an intrinsic that is missing in gcc < 10.

For this target, might trigger another gcc bug on windows that
requires up-to-date gcc 8, 9, or 10, or usage of clang.

Fixes https://github.com/official-stockfish/Stockfish/issues/2975

closes https://github.com/official-stockfish/Stockfish/pull/2976

No functional change

15 months agoMore aligned_alloc changes to support Android
Joost VandeVondele [Mon, 10 Aug 2020 14:14:17 +0000 (16:14 +0200)]
More aligned_alloc changes to support Android

Move to posix_memalign for those platforms, in particular android,
that do not fully support c++17 std::aligned_alloc() (and are not windows)

see https://github.com/official-stockfish/Stockfish/issues/2860

closes https://github.com/official-stockfish/Stockfish/pull/2973

No functional change

15 months agoFix parallel LTO issues on Windows
Guy Vreuls [Tue, 11 Aug 2020 02:38:38 +0000 (04:38 +0200)]
Fix parallel LTO issues on Windows

This adds -save-temps to the linker flags when parallel LTO is used on
MinGW/MSYS.

fixes #2977

closes https://github.com/official-stockfish/Stockfish/pull/2978

No functional change.

15 months agoAdd comments to probCut code
Vizvezdenec [Sat, 18 Jul 2020 13:30:00 +0000 (16:30 +0300)]
Add comments to probCut code

and rename a variable

closes https://github.com/official-stockfish/Stockfish/pull/2819

No functional change

15 months agoAssorted search parameter tune
Unai Corzo [Mon, 10 Aug 2020 13:38:44 +0000 (15:38 +0200)]
Assorted search parameter tune

STC https://tests.stockfishchess.org/tests/view/5f31219090816720665374ec
LLR: 2.96 (-2.94,2.94) {-0.50,1.50}
Total: 3376 W: 487 L: 359 D: 2530
Ptnml(0-2): 17, 253, 1042, 337, 39

LTC https://tests.stockfishchess.org/tests/view/5f3127f79081672066537502
LLR: 2.93 (-2.94,2.94) {0.25,1.75}
Total: 8360 W: 581 L: 475 D: 7304
Ptnml(0-2): 11, 407, 3238, 513, 11

closes https://github.com/official-stockfish/Stockfish/pull/2971

bench: 4733874

15 months agoProvide vectorized NNUE code for SSE2 and MMX targets
Fanael Linithien [Sun, 9 Aug 2020 14:20:45 +0000 (16:20 +0200)]
Provide vectorized NNUE code for SSE2 and MMX targets

This patch allows old x86 CPUs, from AMD K8 (which the x86-64 baseline
targets) all the way down to the Pentium MMX, to benefit from NNUE with
comparable performance hit versus hand-written eval as on more modern
processors.

NPS of the bench with NNUE enabled on a Pentium III 1.13 GHz (using the
MMX code):
  master: 38951
  this patch: 80586

NPS of the bench with NNUE enabled using baseline x86-64 arch, which is
how linux distros are likely to package stockfish, on a modern CPU
(using the SSE2 code):
  master: 882584
  this patch: 1203945

closes https://github.com/official-stockfish/Stockfish/pull/2956

No functional change.

15 months agoCleanup and optimize SSE/AVX code
mstembera [Sun, 9 Aug 2020 23:23:33 +0000 (16:23 -0700)]
Cleanup and optimize SSE/AVX code

AVX512 +4% faster
AVX2 +1% faster
SSSE3 +5% faster

passed non-regression STC:
STC https://tests.stockfishchess.org/tests/view/5f31249f90816720665374f6
LLR: 2.96 (-2.94,2.94) {-1.50,0.50}
Total: 17576 W: 2344 L: 2245 D: 12987
Ptnml(0-2): 127, 1570, 5292, 1675, 124

closes https://github.com/official-stockfish/Stockfish/pull/2962

No functional change

15 months agoMakefile rework/cleanup
sf-x [Sun, 9 Aug 2020 15:01:18 +0000 (18:01 +0300)]
Makefile rework/cleanup

Makefile targets x86-64-sse42, x86-sse3 are removed; x86-64-sse41
is renamed to x86-64-sse41-popcnt (it did enable popcnt).

Makefile variables sse3, sse42, their associated compilation flags
and code in misc.cpp are removed.

closes https://github.com/official-stockfish/Stockfish/pull/2922

No functional change