simple Makefile autodetect for ARCH and BITS - #312
Open
RBEGamer wants to merge 2596 commits into
Open
Conversation
seylab
approved these changes
Jun 9, 2021
chussana-pin
approved these changes
Nov 11, 2021
dav1312
force-pushed
the
master
branch
2 times, most recently
from
October 5, 2023 14:08
9173bb1 to
e0afedd
Compare
Using the already existing pawn_single_push_bb helper which encapsulates the exact same logic. closes #6700 No functional change
[Passed STC](https://tests.stockfishchess.org/tests/live_elo/69cf6a5d1668971c9da23ae1) LLR: 2.93 (-2.94,2.94) <-1.75,0.25> Total: 51104 W: 13376 L: 13180 D: 24548 Ptnml(0-2): 134, 5555, 13963, 5781, 119 [Passed LTC](https://tests.stockfishchess.org/tests/live_elo/69d00a7ee2b443cb2670b5c6) LLR: 2.95 (-2.94,2.94) <-1.75,0.25> Total: 101256 W: 25867 L: 25732 D: 49657 Ptnml(0-2): 59, 10423, 29520, 10576, 50 When processing a promotion we add/remove nonsensical threats, namely, piece threats of pawns on the 1st or 8th rank. This is not only inefficient but also seems to lead to an obscure bug during a promotion–capture + `double_inc_update` that causes a lingering invalid feature. Currently – as an artifact of the training process – these parts of the net are filled with random values in [-1,1], which is why there is a bench change. @ces42 did a helpful analysis: > checking bench with depth 17 actually shows that the bug in has been there all the time since 8e5392d (when threat inputs was merged into master) > i.e. double_inc_update for ThreatFeatureSet has always been slightly bugged After this patch is applied we'll be able to better compress the net by filling the invalid features with 0, w/o changing bench. I expect a 15%-ish size savings. closes #6701 Bench: 2700393
Instead of checking whether a threat index is valid and then writing to a vector if it isn't, we can instead always write to the buffer of the vector, and only increase the size of the vector if the index is valid. This saves some branch mispredictions. passed STC: https://tests.stockfishchess.org/tests/view/69ceb1689f7a7e3fdfc9a44b LLR: 2.95 (-2.94,2.94) <0.00,2.00> Total: 205760 W: 53109 L: 52561 D: 100090 Ptnml(0-2): 603, 22552, 56076, 22992, 657 local speedtest shows ``` Result of 100 runs (cycles) =========================== base (...ockfish.orig) = 296886 +/- 1645 test (...kfish.df0822) = 298862 +/- 1662 speedup % = +0.67 +/- 0.19 [95% CI; t-statistic] p(speedup > 0) = 1.0000 p(speedup > .5%) = 0.9562 CPU: 16 x Intel(R) Core(TM) Ultra 9 185H Hyperthreading: on ``` closes #6703 No functional change
The uci pv suppression and PV roll-back logic in master is a bit convoluted, which makes it hard to reason about the code. In fact, subtle bugs that led to wrong mated-in scores in game play or a mismatch between bestmove and first PV move were only recently fixed. Moreover, in master the uci pv output through `pv()` is called in four different places. This PR proposes to simplify this logic. In this patch, the PV is sent to the GUI from within `iterative_deepening()` only (a) for fail highs/lows or (b) when an iteration for a root move is completed. All other PV outputs are now handled by `start_searching()`, just before the bestmove extraction. This easily ensures that bestmove (and ponder move) will always be in sync with the final PV output. The only noticeable change to master is for multi-threaded searches that do not involve `limits.depth`. Here master would show both the PV from the aborted search in main thread, as well as the new PV from the selected best thread. This patch would only show the latter. While at it, we also remove the requirement to finish at least a depth 1 search and simplify the logic around stalemates/checkmates at root and in particular avoid any PVs that start with `Move::none()`. Passed STC non-reg: LLR: 2.96 (-2.94,2.94) <-1.75,0.25> Total: 72416 W: 18784 L: 18609 D: 35023 Ptnml(0-2): 200, 7762, 20127, 7901, 218 https://tests.stockfishchess.org/tests/view/69d3ef0c33584dad27b3c85a closes #6704 No functional change
Passed STC non-reg: LLR: 2.93 (-2.94,2.94) <-1.75,0.25> Total: 102272 W: 26553 L: 26411 D: 49308 Ptnml(0-2): 269, 11212, 28036, 11346, 273 https://tests.stockfishchess.org/tests/view/69d0f1c5e2b443cb2670b710 closes #6707 No functional change
For some time now we can trust the threads' exact win/loss scores. This allows for some simplification in the best thread selection code. closes #6708 No functional change
Passed non-regression STC: https://tests.stockfishchess.org/tests/view/69d2d3de61a12cebe17ede38 LLR: 2.95 (-2.94,2.94) <-1.75,0.25> Total: 68384 W: 17646 L: 17459 D: 33279 Ptnml(0-2): 200, 8058, 17496, 8231, 207 Passed non-regression LTC: https://tests.stockfishchess.org/tests/view/69d4293433584dad27b3c8cb LLR: 2.95 (-2.94,2.94) <-1.75,0.25> Total: 37290 W: 9582 L: 9380 D: 18328 Ptnml(0-2): 18, 3915, 10579, 4113, 20 closes #6711 Bench: 2923401
the assert checks for `bestThreadDecisive && newThreadDecisive`. However `newThreadDecisive` is not guaranteed at this point, leading to assertion failures. This PR rearranges the conditions to ensure correctness. closes #6716 No functional change
A bit tricky to identify, but in `apple-silicon` builds, specifying a value for `SyzygyPath` could still lead to crashes at certain depths from a SIGBUS/SIGSEGV. This PR replaces the `std::count_if()` adjustment in `tbprobe.cpp`'s `do_probe_table()` with an annotated handwritten loop, avoiding the clang/LTO miscompilation behind the crash. closes #6721 No functional change.
STC: https://tests.stockfishchess.org/tests/view/69d586a64088e069540a22aa LLR: 2.93 (-2.94,2.94) <0.00,2.00> Total: 63520 W: 16579 L: 16233 D: 30708 Ptnml(0-2): 138, 7179, 16825, 7435, 183 LTC: https://tests.stockfishchess.org/tests/view/69d6ddd79ffee997bab13a98 LLR: 2.94 (-2.94,2.94) <0.50,2.50> Total: 227100 W: 58279 L: 57591 D: 111230 Ptnml(0-2): 95, 23435, 65828, 24071, 121 Transition highBestMoveEffort smoothly instead of abruptly. May also be more amenable to future tuning. closes #6729 Bench: 2923401
Trained with the following yaml definition https://github.com/Disservin/nettest/blob/8da0a21503adb341bc89ca20b4e2f166c7c9720d/threats.yaml and the following trainer change official-stockfish/nnue-pytorch@d87c248 Initially I saw a lot of weird 0 scores in the binpack, since then @linrock has explained to me/us their meaning and mentioned that they are actually mostly skipped.. except a very small percentage, however it seems this change was still beneficial. Passed STC: https://tests.stockfishchess.org/tests/view/69db95f43ca80bdf151d4913 LLR: 2.95 (-2.94,2.94) <0.00,2.00> Total: 106272 W: 27916 L: 27489 D: 50867 Ptnml(0-2): 420, 12533, 26841, 12884, 458 Passed LTC: https://tests.stockfishchess.org/tests/view/69ddd9003ca80bdf151d4c4d LLR: 2.94 (-2.94,2.94) <0.50,2.50> Total: 29748 W: 7775 L: 7470 D: 14503 Ptnml(0-2): 24, 3136, 8247, 3445, 22 closes #6730 Bench: 3108996
STC: https://tests.stockfishchess.org/tests/view/69cf29ed1668971c9da23a5a LLR: 3.47 (-2.94,2.94) <-1.75,0.25> Total: 193760 W: 50142 L: 50051 D: 93567 Ptnml(0-2): 534, 21375, 52980, 21448, 543 * Simplify retval and arithmetic semantics * Remove mixing of int and uint8_t I find this overall much easier to reason about (Also rename a DEPTH constant per mstembera's suggestion #5766 (comment)) closes #6702 No functional change
Conditions for no pawn pushes, non-pawn material and last captured piece are already such restrictive that we don't need to artificially set checkers bitboard for legal move generation. Also moving the check for available pawn pushes at first place, as it should already filter about 90% of cases. Passed STC and LTC non-regression tests. LLR: 2.96 (-2.94,2.94) <-1.75,0.25> Total: 171232 W: 44202 L: 44131 D: 82899 Ptnml(0-2): 466, 18809, 47023, 18824, 494 https://tests.stockfishchess.org/tests/view/69d3623b61a12cebe17ededa LLR: 2.94 (-2.94,2.94) <-1.75,0.25> Total: 55662 W: 14424 L: 14247 D: 26991 Ptnml(0-2): 38, 5479, 16609, 5678, 27 https://tests.stockfishchess.org/tests/view/69d64cada07a818b0a25f95e And also neutral on 10k fixed games test with custom stalemate book. Elo: 0.24 ± 0.9 (95%) LOS: 69.9% Total: 10000 W: 4681 L: 4674 D: 645 Ptnml(0-2): 0, 87, 4819, 94, 0 nElo: 1.81 ± 6.8 (95%) PairsRatio: 1.08 https://tests.stockfishchess.org/tests/view/69d7e4183ca80bdf151d43b8 closes #6715 Bench: 2984258
In C++, all standard input streams (including std::istringstream) are initialized with the skipws format flag enabled by default. Because these stream objects are scoped locally and freshly constructed on each loop iteration, we are guaranteed that skipws is already active. closes #6719 No functional change
[Some build environments](https://tests.stockfishchess.org/actions?max_count=400&sort=time&order=desc&action=failed_task&user=gaster319&text=executing) do not have this intrinsic. closes #6722 No functional change.
Use a dedicated struct for keeping track of the pv moves. anematode pointed out that we can use the ValueList for this as well, left that open for a future patch. Passed Non Regression https://tests.stockfishchess.org/tests/view/69d399cfec2dc6794cf18376 LLR: 2.95 (-2.94,2.94) <-1.75,0.25> Total: 267264 W: 68895 L: 68927 D: 129442 Ptnml(0-2): 790, 29238, 73566, 29290, 748 closes #6726 No functional change
Since as of 9417552 the en passant square is only set if there is a legal en passant capture, the special handling is redundant Passed Non-Regression STC: https://tests.stockfishchess.org/tests/view/696ffb6512ee1f6231b96fe8 LLR: 2.95 (-2.94,2.94) <-1.75,0.25> Total: 79072 W: 20380 L: 20210 D: 38482. closes #6560 No functional change
We maintain support for quite a few x86 ISA extensions but relied on the end
user to select the binary that is best for their system. We can detect at
runtime which architecture to use, and ship a single binary (per OS/ISA
combination). This PR does so, maintaining performance.
This is what I've landed on after quite a few iterations. Basically, we build
each arch separately, use `cpuid` to select one, and jump to that arch's main
function. Some details:
- The preprocessor macro `UNIVERSAL_BINARY` is defined when building for the universal binary.
- The Makefile target `universal-object-[no]pgo` is added, which produces a `stockfish.o` in each arch's build directory.
- To prevent symbol collisions between the multiple builds, we `#define Stockfish Stockfish_[arch]`. Furthermore we wrap the `main` function in `namespace Stockfish { }`.
- We can still get PGO data by linking to a per-arch binary, with `Stockfish_x86_64_avx2::main` as `main`, and running `bench`.
- For arches not supported by the host we can use Intel SDE – this is what we do in CI.
- When embedding the NNUE, we use C++26/C23 `#embed` instead of `INCBIN`. The issue with `INCBIN` is that it injects assembly directives that we have no control over.
- To ensure there's only one copy of the networks in the final binary, we define the network data as weak symbols in each per-arch build. Then, in the final link, we add a special nnue_embed.cpp which provides strong symbols.
- This does require GCC 15. Statically linking libstdc++ allows the binary to still run on older OSes though. And latest msys2 already does static linking + is based off GCC 15.2.0
building can be as simple as
make -j profile-build ARCH=x86-64-universal
or using the sde if the host doesn't support all architectures supported in the
universal binary, and the default host compiler is not recent enough.
make -j profile-build ARCH=x86-64-universal RUN_PREFIX="/path/to/sde -future --" CXX=g++-15
This change is also integrated in CI, so universal binaries are available as
downloadable artifacts.
Next we could also do ARM64, especially Android (all Apple silicon users use
the same binary). It also might be worth getting this to work with clang.
closes #6740
No functional change
Updates the information printed in the "eval" command. Removes piece value estimation, and adds both small and big net eval information. closes #6743 No functional change
Passed STC: LLR: 2.97 (-2.94,2.94) <0.00,2.00> Total: 677344 W: 175301 L: 174041 D: 328002 Ptnml(0-2): 2018, 79669, 174166, 80673, 2146 https://tests.stockfishchess.org/tests/view/69cfae021668971c9da23b80 Passed LTC: LLR: 2.95 (-2.94,2.94) <0.50,2.50> Total: 46470 W: 11988 L: 11651 D: 22831 Ptnml(0-2): 37, 4877, 13055, 5244, 22 https://tests.stockfishchess.org/tests/view/69dcd13c3ca80bdf151d4ab5 Passed VLTC (non-reg.): LLR: 2.95 (-2.94,2.94) <-1.75,0.25> Total: 58986 W: 15120 L: 14947 D: 28919 Ptnml(0-2): 7, 5835, 17636, 6008, 7 https://tests.stockfishchess.org/tests/view/69debc6c5c67bee7d241ffe9 closes #6741 Bench: 2723949
closes #6735 No functional change
Fix MultiArray::at() which was marked noexcept despite delegating to std::array::at() that throws on out-of-range. Per review feedback, at() now keeps standard bounds-checking semantics (only noexcept removed), and operator[] gains an assert for debug-mode safety. All eight correction history lookups in search.cpp (correction_value and update_correction_history) are switched from .at(us) to [us]. Passed STC: https://tests.stockfishchess.org/tests/view/69de7a735c67bee7d241ff92 LLR: 2.96 (-2.94,2.94) <-1.75,0.25> Total: 77536 W: 20031 L: 19860 D: 37645 Ptnml(0-2): 221, 8411, 21338, 8572, 226 closes #6734 No functional change
to a simpler and cleaner version Passed non-reg test: LLR: 2.95 (-2.94,2.94) <-1.75,0.25> Total: 175648 W: 45156 L: 45091 D: 85401 Ptnml(0-2): 459, 19228, 48383, 19297, 457 https://tests.stockfishchess.org/tests/view/69d8d9263ca80bdf151d454a closes #6733 No functional change
In setup_benchmark, the variables evaluating the current ply simply mirror the exact iterative progress of the string arrays loop (since ply just increments by 1 with every pass). We can delete the separate integer assignment and execute it securely within the loop declaration. closes #6732 No functional change
closes #6744 No functional change
The tail byte was read as signed char, so bytes >= 0x80 sign-extended and erased previously accumulated bytes in the hash. Read it as u8. closes #7067 No functional change Co-authored-by: ppigazzini <pasquale.pigazzini@gmail.com>
elapsed() returns nodes while nodestime is active, but check_time compared it against limits.movetime in milliseconds without conversion, so movetime silently meant nodes. Convert movetime to nodes before init's early return. closes #7068 No functional change Co-authored-by: ppigazzini <pasquale.pigazzini@gmail.com>
Worker::clear partitioned correctionHistory and pawnHistory by NUMA thread index, but then filled the shared continuationHistory from every worker. Since continuationHistory is constant-size and shared only within a NUMA node, have only thread 0 of each node initialize it closes #7069 No functional change Co-authored-by: ppigazzini <pasquale.pigazzini@gmail.com>
speedtest multiplied user-supplied seconds and threads in int, which could overflow and produce negative movetime or Hash values. Clamp each input to the range the option layer accepts and report corrections. closes #7070 No functional change Co-authored-by: ppigazzini <pasquale.pigazzini@gmail.com>
STC Ice Lake: https://tests.stockfishchess.org/tests/view/6a7d0c325ebead9b2e78724a LLR: 4.20 (-2.94,2.94) <-1.75,0.25> Total: 266880 W: 67650 L: 67577 D: 131653 Ptnml(0-2): 373, 28574, 75459, 28675, 359 STC Non Ice Lake: https://tests.stockfishchess.org/tests/view/6a7d6a8f5ebead9b2e7872d0 LLR: 3.12 (-2.94,2.94) <-1.75,0.25> Total: 278112 W: 71067 L: 71095 D: 135950 Ptnml(0-2): 485, 29231, 79629, 29249, 462 This is a fixed version of #7048 which had a regression on Ice Lake. Thanks to @Torom for catching it and to @anematode for suggesting separate tests. closes #7075 No functional change
Make the existing MADV_HUGEPAGE request effective for the Linux shared NNUE mapping. Map large shared objects at a 2 MiB-aligned address and request MADV_COLLAPSE after the creating process initializes the network. pyshbench, 30 runs: ```text Platform / target Speedup 95% CI Graviton 3 / armv8-dotprod +2.675% [+2.495%, +2.856%] AMD Zen 3 / x86-64-bmi2 +0.479% [+0.182%, +0.776%] Intel Sapphire / x86-64-avx512icl +2.073% [+1.675%, +2.470%] ``` STC passed: ```text LLR: 2.96 (-2.94,2.94) <0.00,2.00> Total: 37344 W: 9716 L: 9419 D: 18209 Ptnml(0-2): 54, 3890, 10497, 4167, 64 ``` https://tests.stockfishchess.org/tests/view/6a80cc2adb2c105c37ecd9cf closes #7058 No functional change
Update the release pipeline so that we can create immutable releases. The problem of the previous pipeline was that the release was created first and only afterwards the binaries were added, which is not possible with immutable releases. https://github.com/Disservin/Stockfish/actions/runs/32180223600 <img width="1113" height="710" alt="image" src="https://github.com/user-attachments/assets/1e4945f7-3557-4c1f-afe2-7be0ed2b1060" /> closes #7076 No functional change
pthread_create's return value was ignored, so a failed spawn left the engine hanging forever waiting for a searching flag no thread would clear. Report the error and exit. Avoid using std::thread. Intends to consolidate #7060 and #7065 by making `NativeThread` more customizable. closes #7072 No functional change Co-authored-by: Timothy Herchen <timothy.herchen@gmail.com> Co-authored-by: ppigazzini <pasquale.pigazzini@gmail.com>
Per Rob's comment https://discord.com/channels/435943710472011776/813919248455827515/1539144885255536761 and in the spirit of similar PRs. Feel free to comment whatever uncontroversial claenups and I will apply them Changes: - Add Jonathan to `AUTHORS` which we missed when merging pawn-pair inputs - Mr. Dubslow's changes from #6942 - mstembera's style tweaks - niklasf's CLOEXEC enjoyment - lemtea's `.size() == 0` -> `empty()` closes #7074 No functional change Co-authored-by: Dubslow <bunslow@gmail.com>
Bug: On MSVC sf_always_inline expands to ```__forceinline```, which already implies ```inline```, so definitions writing both triggered "warning C4141 ('inline' used more than once)".
Fix: Moved ```inline``` into the macro and removed it from all definition sites.
GCC/Clang still get ```inline __attribute__((always_inline))```. ```inline``` is in the macro since the attribute alone would not set linkage.
[Edit]
Also adds MSVC coverage to CI as requested.
closes #7078
No functional change
Failure-path tests were masking valgrind leak reports because --error-exitcode=42 was being treated as the expected non-zero exit code. Add an expect_failure flag so these tests run valgrind with --leak-check=no, while still checking ERROR SUMMARY for real memory errors. Populate CLI output from stdout/stderr so valgrind output is visible to postfix_check, and only print the captured test output when a test fails. closes #7081 No functional change
This PR simplifies the recent patch #7040, while still allowing Stockfish to consistently pass the CI `go mate` run involving 61 positions with DTM from 1 to 4 plies. Results of a deterministic th1 run covering these 61 positions are as follows. ``` Loaded 61 FENs with 61 bm values, with |bm| (min avg max): 1 2 2. Matetrack started for ./sf.master on matetrack.epd matedtrack.epd with --bmMax 2 --mate 0 ... 100%|███████████████████████████████████████████| 61/61 [00:17<00:00, 3.54it/s] Using ./sf.master on matetrack.epd matedtrack.epd with --bmMax 2 --mate 0 Engine ID: Stockfish dev-20260819-229f6339 Total FENs: 61 Found mates: 61 Best mates: 61 Best mate statistics: |bm| = 1 - mates found: 21 = 100.0% of 21; nodes (min avg max): 4 343 1372, depth (min avg max): 1 4 12 |bm| = 2 - mates found: 40 = 100.0% of 40; nodes (min avg max): 16 402441 15454156, depth (min avg max): 1 10 24 All best mates found: 61 = 100.0% of 61; nodes (min avg max): 4 264014 15454156, depth (min avg max): 1 8 24 ``` ``` Loaded 61 FENs with 61 bm values, with |bm| (min avg max): 1 2 2. Matetrack started for ./sf.patch on matetrack.epd matedtrack.epd with --bmMax 2 --mate 0 ... 100%|███████████████████████████████████████████| 61/61 [00:02<00:00, 24.64it/s] Using ./sf.patch on matetrack.epd matedtrack.epd with --bmMax 2 --mate 0 Engine ID: Stockfish dev-20260824-1a3ceee9 Total FENs: 61 Found mates: 61 Best mates: 61 Best mate statistics: |bm| = 1 - mates found: 21 = 100.0% of 21; nodes (min avg max): 4 343 1372, depth (min avg max): 1 4 12 |bm| = 2 - mates found: 40 = 100.0% of 40; nodes (min avg max): 16 20068 255722, depth (min avg max): 1 9 21 All best mates found: 61 = 100.0% of 61; nodes (min avg max): 4 13277 255722, depth (min avg max): 1 8 21 ``` The above shows that patch outperforms master 2s:17s, and 13277:264014 nodes. In addition, patch shows a healthy increase in (best) mates found on matetrack: ``` Using ./sf.patch on matetrack.epd with --nodes 1000000 Engine ID: Stockfish dev-20260824-1a3ceee9 Total FENs: 6554 Found mates: 3700 Best mates: 2457 ``` This compares with 3202, 2263 for master, see https://github.com/vondele/matetrack/blob/c88991a2558b326e2119ca0e5cc9554d60e559ab/matetrack1000000.csv#L4530. The numbers for the classic mate suite are: ``` Loaded 280 FENs with 280 bm values, with |bm| (min avg max): 10 15 21. Matetrack started for ./sf.patch on classic280.epd with --nodes 1000000 ... 100%|███████████████████████████████████████████| 56/56 [00:26<00:00, 2.13it/s] Using ./sf.patch on classic280.epd with --nodes 1000000 Engine ID: Stockfish dev-20260824-1a3ceee9 Total FENs: 280 Found mates: 128 Best mates: 31 ``` This compares with 165, 24 for master, see https://github.com/vondele/matetrack/blob/c88991a2558b326e2119ca0e5cc9554d60e559ab/classic1000000.csv#L4530. Passed STC non-reg: LLR: 2.93 (-2.94,2.94) <-1.75,0.25> Total: 41504 W: 10764 L: 10559 D: 20181 Ptnml(0-2): 61, 4620, 11196, 4803, 72 https://tests.stockfishchess.org/tests/view/6a85b5af9960adfadf924292 Passed LTC non-reg: LLR: 2.94 (-2.94,2.94) <-1.75,0.25> Total: 111234 W: 28827 L: 28702 D: 53705 Ptnml(0-2): 32, 11680, 32064, 11813, 28 https://tests.stockfishchess.org/tests/view/6a86ef485b1b38ebda864e69 closes #7089 Bench: 2502027
This PR fixes two minor issues with master's PV extension code. 1. When 'nodestime' is active, master may abort PV extension if it takes too long (as measured in ms), rendering the produced results nondeterministic. The patch always completes the PV extensions. This has no effect on playing strength because the local calls to `pos.do_move()` do not update the global used nodes counter. 2. If standard time management is active, and the game is played with `multiPV > 1`, then extending several PVs may repeatedly use up to `moveOverhead / 2` ms, possibly leading to time losses. The patch fixes this by only extending the PV of the best move during game play. closes #7079 No functional change
Master branch does not compile on some Unix systems since commit d077f9a (for instance compilation fails on Mac OS X 10.13.6 using GNUC 10.4.2). The problem is that on such systems CMSG_SPACE() is not a compile time constant, so it cannot be used to allocate the size of the message buffer we use in shm_unix.h The fix is to use a dynamic array instead of a stack array for the buffer. We use a standard C++ vector to get automatic memory management, and align it with our utilities in memory.h. fixes #7092 closes #7093 No functional change
should fix the common case of downloads not being successful. Similarly cache the syzygy download. closes #7094 No functional change Co-authored-by: Joost VandeVondele <joost.vandevondele@gmail.com>
Worker/SharedHistories allocations on Linux were served by glibc malloc arenas, which survive thread destruction and can be reused by threads bound to a different NUMA node. This caused non-deterministic remote memory and lower NPS when Threads was set multiple times. Route aligned_large_pages_alloc_with_hint() through direct mmap() on Linux, storing the base/size so aligned_large_pages_free() can munmap() the original mapping. Merge the existing x86_64 MAP_HUGETLB bookkeeping into the same registry. fixes #6516 Benchmark: 288 threads, 36864 MB hash, depth 15 (10 runs per config) ``` | Build | Case | Avg NPS | Min NPS | Max NPS | Range | |--------|--------------------------|-------------|-------------|-------------|-----------| | master | Case1 (Threads set x2) | 253,196,264 | 252,034,443 | 254,396,860 | 2,362,417 | | master | Case2 (Threads set x1) | 248,224,546 | 247,304,918 | 249,306,273 | 2,001,355 | | fix | Case1 (Threads set x2) | 254,675,607 | 253,991,054 | 255,600,403 | 1,609,349 | | fix | Case2 (Threads set x1) | 254,931,057 | 253,979,373 | 256,334,683 | 2,355,310 | ``` closes #7095 No functional change
update to current closes #7098 No functional change
Simplify the optimism scaling formula, keeping the i64 cast for safety. LLR: 2.93 (-2.94,2.94) <-1.75,0.25> Total: 506336 W: 129447 L: 129738 D: 247151 Ptnml(0-2): 960, 59384, 132718, 59199, 907 Passed STC non-reg: https://tests.stockfishchess.org/tests/view/6a79ff99a7e815d01b5603ad Passed LTC non-reg: LLR: 2.94 (-2.94,2.94) <-1.75,0.25> Total: 138276 W: 35651 L: 35552 D: 67073 Ptnml(0-2): 37, 14851, 39261, 14954, 35 https://tests.stockfishchess.org/tests/view/6a842ae29960adfadf924008 closes #7097 Bench: 2497913
Clean up and add some comments. closes #7100 No functional change
passed STC: https://tests.stockfishchess.org/tests/view/6a8d043108c0f6a39c9e7eba LLR: 2.94 (-2.94,2.94) <-1.75,0.25> Total: 289440 W: 73563 L: 73617 D: 142260 Ptnml(0-2): 356, 30896, 82225, 30932, 311 closes #7090 No functional change
Passed STC https://tests.stockfishchess.org/tests/view/6a8ab17b08c0f6a39c9e79ae LLR: 2.96 (-2.94,2.94) <0.00,2.00> Total: 96192 W: 24637 L: 24264 D: 47291 Ptnml(0-2): 136, 9961, 27552, 10288, 159 Passed avx2 STC: https://tests.stockfishchess.org/tests/view/6a8cd16708c0f6a39c9e7e59 LLR: 2.94 (-2.94,2.94) <0.00,2.00> Total: 61024 W: 15789 L: 15457 D: 29778 Ptnml(0-2): 108, 6534, 16905, 6848, 117 Failed avx512icl STC: https://tests.stockfishchess.org/tests/view/6a8cd14d08c0f6a39c9e7e57 LLR: -2.96 (-2.94,2.94) <0.00,2.00> Total: 90368 W: 22968 L: 23088 D: 44312 Ptnml(0-2): 132, 9707, 25587, 9665, 93 Context: #7051 (comment) Credit to bobbypaper for inspiring the idea. Gated to AVX2 because it seems neutral at best on other architectures. Locally, despite the avx512icl failure: Result of 500 runs base (...kfish.master) = 1907788 +/- 1384 test (./stockfish ) = 1930017 +/- 1426 diff = +22230 +/- 1978 speedup = +0.0117 P(speedup > 0) = 1.0000 closes #7096 No functional change
Fixes losses on time in cyclic time controls. Results of master vs fix (40/1, 1t, 64MB, UHO_Lichess_4852_v1.epd): Elo: -48.65 +/- 9.84, nElo: -76.63 +/- 15.29 LOS: 0.00 %, DrawRatio: 40.83 %, PairsRatio: 0.48 Games: 1984, Wins: 451, Losses: 727, Draws: 806, Points: 854.0 (43.04 %) Ptnml(0-2): [81, 316, 405, 178, 12], WL/DD Ratio: 1.60 Master timeouts: 557, Fix timeouts: 0 Results of master vs fix (40/10, 1t, 64MB, UHO_Lichess_4852_v1.epd): Elo: -8.10 +/- 3.45, nElo: -15.98 +/- 6.81 LOS: 0.00 %, DrawRatio: 52.00 %, PairsRatio: 0.84 Games: 10000, Wins: 2463, Losses: 2696, Draws: 4841, Points: 4883.5 (48.84 %) Ptnml(0-2): [39, 1268, 2600, 1073, 20], WL/DD Ratio: 1.08 Master timeouts: 116, Fix timeouts: 0 closes #7117 No functional change
This PR updates the internal WDL model, using data from 2M games played by the revisions since 229f633. The normalizing constant decreases significantly from 385 to 304. That means the current 78cp will roughly map to 100cp. ``` > ./updateWDL.sh --firstrev 229f633 Running: ./updateWDL.sh --firstrev 229f633 --lasttrev HEAD --materialMin 17 --EloDiffMax 5 started at: Tue 1 Sep 14:31:12 CEST 2026 Look recursively in directory pgns for games with max nElo difference 5 using books matching "UHO_Lichess_4852_v..epd" for SF revisions between 229f633 (from 2026-08-19 18:48:49 +0200) and HEAD (from 2026-08-29 09:16:46 +0200). Based on 109147283 positions from 2046412 games, NormalizeToPawnValue should change from 385 to 304. In the updated WDL model, 78cp would roughly map to 100cp. ended at: Tue 1 Sep 14:32:20 CEST 2026 ``` ``` > cat scoreWDL.log Converting evals with NormalizeData = {'momType': 'material', 'momMin': 17, 'momMax': 78, 'momTarget': 58, 'as': [-72.32565836, 185.93832038, -144.58862193, 416.44950446]}. Reading eval stats from updateWDL.json. Retained (W,D,L) = (26673519, 54948054, 27525710) positions. Saved distribution plot to updateWDLdistro.png. Fit WDL model based on material. Initial objective function: 0.3550415621263458 Final objective function: 0.3550400701834142 Optimization terminated successfully. const int NormalizeToPawnValue = 304; Corresponding spread = 61; Corresponding normalized spread = 0.19927805405945914; Draw rate at 0.0 eval at material 58 = 0.986853004859315; Parameters in internal value units: p_a = ((-142.721 * x / 58 + 372.352) * x / 58 + -340.711) * x / 58 + 415.235 p_b = ((5.938 * x / 58 + 15.613) * x / 58 + -30.578) * x / 58 + 69.639 constexpr double as[] = {-142.72052667, 372.35176398, -340.71073572, 415.23490212}; constexpr double bs[] = {5.93832785, 15.61267078, -30.57816876, 69.63866711}; Preparing plots. Saved graphics to updateWDL.png. Total elapsed time = 30.92s. ``` The patch only affects the displayed `cp` and `wdl` values, and nothing else. closes #7120 No functional change
Repro with `-ftrapv`: ``` position fen 4rk2/2r2p2/3R1Bpn/7p/1pPp3P/1P3P2/3p2P1/3R3K w - - 0 36 go ``` ``` info depth 27 seldepth 61 multipv 1 score cp 169 nodes 5778888 nps 514090 hashfull 964 tbhits 0 time 11241 pv d1d2 h6f5 d6b6 c7d7 h1g1 f8g8 d2e2 e8a8 e2a2 a 8c8 g1f2 g8h7 f6g5 h7g7 a2e2 c8a8 e2d2 a8e8 g5f6 g7h7 b6b4 e8e6 f6g5 f7f6 g5f4 f5h4 c4c5 g6g5 f4h2 h4f5 b4b6 e6b6 c5b6 f5e7 h2c7 e7c6 f2e2 d7d5 info depth 28 seldepth 62 multipv 1 score cp 169 nodes 6673438 nps 507833 hashfull 984 tbhits 0 time 13141 pv d1d2 h6f5 d6b6 c7d7 h1g1 f8g8 d2e2 e8a8 e2a2 a 8c8 g1f2 g8h7 f6g5 h7g7 a2d2 c8e8 g5f6 g7h7 b6b4 e8e6 f6g5 f7f6 g5f4 e6c6 g2g3 h7g8 b4b5 d4d3 c4c5 f5d4 d2d3 d4b5 d3d7 c6c5 f4d2 c5c2 f2e3 b5c3 Thread 4 "stockfish" received signal SIGABRT, Aborted. (gdb) bt [snip] #5 0x00007ffff7bd8e55 in __mulvsi3 () from /lib/x86_64-linux-gnu/libgcc_s.so.1 #6 0x000000000043b8ce in Stockfish::(anonymous namespace)::update_continuation_histories(Stockfish::Search::Stack*, Stockfish::Piece, Stockfish::Square, in t) () ``` After this fix, the worst-case near-overflows across Stockfish that I could find (which are still this one!) still have a healthy margin: ``` MUL closest to overflow, margin: 1086513607 (10300680 x 103 -> 1060970040) ``` closes #7123 No functional change
Cast to `unsigned char` before calling these functions, matching the pattern already used elsewhere in the codebase (`UCIEngine::to_lower`, `CaseInsensitiveLess::operator()`). closes #7126 No functional change Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
closes #7127 No functional change
When `availableNodes` hits 0 in master, on the next move we hit UB in `TimeManagement::init()`. More generally, the UB can happen whenever `availableNodes < npmsec`. That is because `limits.time[us] == 0` gets passed into `scaledTime`, which is then used in `log10` to define `logTimeInSec`, leading to `-inf` in `optConstant` and hence `+inf` in `optScale`, if `originalTimeAdjust < 0`. Casting that `+inf` double value to the integer valued `optimumTime` via `TimePoint(std::max(1.0, optScale * timeLeft))` is undefined behavior. The patch clamps `scaledTime` from below to 1. closes #7129 No functional change
- Add a comment for each function of misc.cpp - Improve comment indentation in tt.h and tt.cpp closes #7130 No functional change
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
20 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
I added a simple autodetect flag for the makefile:
make build ARCH=autodetect### 2.1.1 Simple Defaults Autodetection ifeq ($(ARCH),autodetect) arch = $(shell uname -m) bits = $(shell getconf LONG_BIT) endifI used the modification in my docker build system.
Maybe its useful!