Improves throughput by summing 2 intermediate dot products using 16 bit addition before upconverting to 32 bit.
Potential saturation is detected and the code-path is avoided in this case.
The saturation can't happen with the current nets,
but nets can be constructed that trigger this check.
STC https://tests.stockfishchess.org/tests/view/5fd40a861ac1691201888479
LLR: 2.94 (-2.94,2.94) {-0.25,1.25}
Total: 25544 W: 2451 L: 2296 D: 20797
Ptnml(0-2): 92, 1761, 8925, 1888, 106
about 5% speedup
closes https://github.com/official-stockfish/Stockfish/pull/3261
No functional change
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.
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
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
despite usage of alignas, the generated (avx2/avx512) code with older compilers needs to use
unaligned loads with older gcc (e.g. confirmed crash with gcc 7.3/mingw on abrok).
Better performance thus requires gcc >= 9 on hardware supporting avx2/avx512
closes https://github.com/official-stockfish/Stockfish/pull/2969
No functional change
Remove temporary array of shelters and avoid iterating over it each time to find
if the shelter values after castling are better than the current value.
Work done on top of https://github.com/official-stockfish/Stockfish/pull/2277
Speed benchmark did not measure any difference.
No functional change
We don't need to pass the king square as an explicit parameter to the functions
king_safety() and do_king_safety() since we already pass in the position.
STC:
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 69686 W: 14894 L: 14866 D: 39926
http://tests.stockfishchess.org/tests/view/5be84ac20ebc595e0ae3283c
No functional change.
Compiling the current master with MSVC gives the following error:
```
search.cpp(956): error C2660: 'operator *': function does not take 1 arguments
types.h(303): note: see declaration of 'operator *'
```
This was introduced in commit:
https://github.com/official-stockfish/Stockfish/commit/88de112b84a5285c2afb3e075a05c2ab8ad3fd33
We use a suggestion by @vondele to fix the error, thanks!
No functional change.
Default template parameters values and recursive functions do not play well
together. Fix for below errors that showed up after updating to latest MSVC.
````
tbprobe.cpp(1156): error C2672:
'search': no matching overloaded function found
tbprobe.cpp(1198): error C2783:
'Tablebases::WDLScore `anonymous-namespace'::search(Position &,Tablebases::ProbeState *)':
could not deduce template argument for 'CheckZeroingMoves'
````
Closes https://github.com/official-stockfish/Stockfish/pull/1594
No functional change.
Simplification: in king danger, include all blockers and not only pinned
pieces, since blockers enemy pieces can result in discovered checks which
are also bad.
STC http://tests.stockfishchess.org/tests/view/5af35f9f0ebc5968e6523fe9
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 145781 W: 29368 L: 29478 D: 86935
LTC http://tests.stockfishchess.org/tests/view/5af3cb430ebc5968e652401f
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 76398 W: 11272 L: 11232 D: 53894
I also incorrectly scheduled STC with [0,5] which it failed.
http://tests.stockfishchess.org/tests/view/5af283c00ebc5968e6523f33
LLR: -2.94 (-2.94,2.94) [0.00,5.00]
Total: 12338 W: 2451 L: 2522 D: 7365
Closes https://github.com/official-stockfish/Stockfish/pull/1593
bench: 4698290
----------------------------------------
Thanks to @vondele and @Rocky640 for a cleaner version of the patch,
and the following comments!
> Most of the pinned, (or for this pull request, blocking) squares were
> already computed in the unsafeChecks, the only missing squares being:
>
> a) squares attacked by a Queen which are occupied by friendly piece
> or "unsafe". Note that adding such squares never passed SPRT[0,5].
>
> b) squares not in mobilityArea[Us].
>
> There is a strong relationship between the blockers and the unsafeChecks,
> but the bitboard unsafeChecks is still useful when the checker is not
> aligned with the king, and the checking square is occupied by friendly
> piece or is "unsafe". This is always the case for the Knight.
1) Use make_bitboard() in Bitboards::init()
2) Fix MSVC warning: search.h(85): warning C4244: '=': conversion from
'TimePoint' to 'int', possible loss of data.
Closes https://github.com/official-stockfish/Stockfish/pull/1524
No functional change.
Make magic_index() a member of Magic since it uses all it's members
and keep us from having to pass the function pointer around to
init_magics().
No functional change
Closes#1146
Gather all magic relevant data into a struct.
This changes memory layout putting everything necessary for processing a single square
in the same memory location thus speeding up access.
Original patch by @snicolet
No functional change.
Closes#1127Closes#1128
This eliminates alignment padding and reduces size from 48 to 40 bytes.
This makes the material HashTable smaller and more cache friendly.
No functional change
Closes#1013
Also a speedup(about 1%) on 64-bit w/o hardware popcnt
Retire Max15 and Full template parameters
(Contributed by Marco Costalba)
Now that we have just SW and HW versions, use
template default parameter to get rid of explicit
template parameters.
Retire bitcount.h and move the only defined
function to bitboard.h
No functional change
Resolves#620
Fix a bug where we could stop the search after only 10% of time used due to a matching easy move but later switch to a different move that was never pre-screened as easy due to SMP thread select.
STC:
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 27227 W: 4910 L: 4800 D: 17517
LTC:
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 40368 W: 5826 L: 5733 D: 28809
Resolves#521
This is used by std::stable_sort() to sort moves from highest score to lowest score.
1) The comment is incorrect since highest to lowest means descending.
2) It's more natural to implement a less operator using another less operator rather than a greater operator.
No functional change.
Resolves#504
Only refresh TT entry when it's really necessary.
This should give a small speed boost for some machines.
And it's a risk-free change.
No functional change.
Resolves#429
v = value without ep capture being considered
v1 = value of the ep capture
The correct logic is:
if without e.p. capture we are losing, and the value of e.p is either draw, or win or "loss, but 50 move rule saves us", then we should use the value of ep capture.
Credit and thanks to syzygy1 and lantonov !
No functional change (except with syzygy bases)
Resolves#415Resolves#394
Spend much less time in positions where one move is much better than all other alternatives.
We carry forward pv stability information from the previous search to identify such positions.
It's based on my old InstaMove idea but with two significant improvements.
1) Much better instability detection inside the search itself.
2) When it's time to make a FastMove we no longer make it instantly but still spend at least 10% of normal time verifying it.
Credit to Gull for the inspiration.
BIG thanks to Gary because this would not work without accurate PV!
20K
ELO: 8.22 +-3.0 (95%) LOS: 100.0%
Total: 20000 W: 4203 L: 3730 D: 12067
STC
LLR: 2.96 (-2.94,2.94) [-1.50,4.50]
Total: 23266 W: 4662 L: 4492 D: 14112
LTC
LLR: 2.95 (-2.94,2.94) [0.00,6.00]
Total: 12470 W: 2091 L: 1931 D: 8448
Resolves#283
I went through all the individual compile options that differ between
-fprofile-generate/-fprofile-use and -fprofile-arcs/-fbranch-probabilities
and distilled the speed difference down to only turning off
-fno-peel-loops and -fno-tracer. Using this we still get the full speedup
(maybe a bit more because other optimizations stay on) and it's also much cleaner
because we can get rid of the "@rm -f ucioption.gc*" hack for all versions of gcc.
No functional change.
Resolves#237
I have noticed that increasing the bench depth produces
progressively smaller and slightly faster executables at
the cost of longer compile times. Also using bench "time"
instead of "depth" seems to produce slightly smaller/faster
executables given comparable compile times.
I have made a new Makefile that generates smaller and
about 1% to 2% faster profile executables at only a
little extra compile time. On my mobile 2GHz i7 a
full profile build time goes from 3'48" to 4'13" and
the exe goes down by 5% from 416,310 bytes to 395,567
bytes.
No functional change.
