Files
stockfish/src/evaluate.cpp
T
69a01b88f3 Use shared memory for network weights
This enables different Stockfish processes that use the same weights to use the
same memory. The approach establishes equivalence by memory content, and is
compatible with NUMA replication.  The benefit of sharing is reduced memory usage
and a speedup thanks to improved (inter-process) caching of the network in the
CPUs cache, and thus reduced bandwidth usage to main memory. Even though this
change doesn't benefit a user running a single process, this helps on fishtest
or e.g.  for Lichess, when multiple games run concurrently, or multiple
positions are analyzed in parallel.

This concept was probably first introduced in the Monty engine
(https://github.com/official-monty/Monty/pull/62), after a discussion in
https://github.com/official-stockfish/fishtest/issues/2077 on the issue of
memory pressure. Measurements based on Torch
(https://github.com/user-attachments/files/21386224/verbatim.pdf) further
suggested that large gains were possible. Multiple other engines have
adopted this 'verbatim' format as well.

The implementation here adds the flexibility needed for SF, for example, retains
the ability to bundle compressed networks with the binary, to load nets by uci
option, and to distribute the shared nets to the proper NUMA region. This
flexibility comes with a fair amount of complexity in the implementation, such
as OS specific code, and fallback code.

For most users this should be transparent. However, for example, those running
docker containers should ensure the `--ipc` flag is set correctly, and
`--shm-size` is sufficiently large.

The benefits of this patch significantly depend on hardware, with systems with
many cores and a large (O(150MB), the net size) L3 cache benefitting typically
most.  On such systems SF speedups (as measured via nps playing games with
large concurrency but just 1 thread) can be 38%, which results in master vs.
patch Elo which gains about 25 Elo.

```
   # PLAYER             :  RATING  ERROR   POINTS  PLAYED   (%)
   1 shared_memoryPR    :    24.8    1.9  39432.0   73728    53
   2 master             :     0.0   ----  34296.0   73728    47
```

In a multithreaded setup, where weights are already shared, that benefit is smaller,
for example on the same HW as above, but with 8t for each side.
```
   # PLAYER             :  RATING  ERROR  POINTS  PLAYED   (%)
   1 shared_memoryPR    :     5.2    3.5  9351.0   18432    51
   2 master             :     0.0   ----  9081.0   18432    49
```

On fishtest with a typical hardware mix of our contributors, the following was measured:

STC, 60k games
https://tests.stockfishchess.org/tests/view/69074a49ea4b268f1fac236c
Elo: 4.69 ± 1.4 (95%) LOS: 100.0%
Total: 60000 W: 16085 L: 15275 D: 28640
Ptnml(0-2): 154, 6440, 16053, 7148, 205
nElo: 9.38 ± 2.8 (95%) PairsRatio: 1.12

To verify correctness with a single process on a NUMA architecture,
speedtest was used, confirming near equivalence:
```
master:        Average (over 10):  296236186
shared_memory: Average (over 10):  295769332
```
Currently, using large pages for the shared network weights is not always possible,
which can lead to a small slowdown (1-2%), in case a single process is run.

closes https://github.com/official-stockfish/Stockfish/pull/6173

No functional change

Co-authored-by: disservin <disservin.social@gmail.com>
Co-authored-by: Joost VandeVondele <Joost.VandeVondele@gmail.com>
2025-11-02 16:04:09 +01:00

125 lines
4.8 KiB
C++

/*
Stockfish, a UCI chess playing engine derived from Glaurung 2.1
Copyright (C) 2004-2025 The Stockfish developers (see AUTHORS file)
Stockfish is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
Stockfish is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see <http://www.gnu.org/licenses/>.
*/
#include "evaluate.h"
#include <algorithm>
#include <cassert>
#include <cmath>
#include <cstdlib>
#include <iomanip>
#include <iostream>
#include <memory>
#include <sstream>
#include <tuple>
#include "nnue/network.h"
#include "nnue/nnue_misc.h"
#include "position.h"
#include "types.h"
#include "uci.h"
#include "nnue/nnue_accumulator.h"
namespace Stockfish {
// Returns a static, purely materialistic evaluation of the position from
// the point of view of the side to move. It can be divided by PawnValue to get
// an approximation of the material advantage on the board in terms of pawns.
int Eval::simple_eval(const Position& pos) {
Color c = pos.side_to_move();
return PawnValue * (pos.count<PAWN>(c) - pos.count<PAWN>(~c))
+ (pos.non_pawn_material(c) - pos.non_pawn_material(~c));
}
bool Eval::use_smallnet(const Position& pos) { return std::abs(simple_eval(pos)) > 962; }
// Evaluate is the evaluator for the outer world. It returns a static evaluation
// of the position from the point of view of the side to move.
Value Eval::evaluate(const Eval::NNUE::Networks& networks,
const Position& pos,
Eval::NNUE::AccumulatorStack& accumulators,
Eval::NNUE::AccumulatorCaches& caches,
int optimism) {
assert(!pos.checkers());
bool smallNet = use_smallnet(pos);
auto [psqt, positional] = smallNet ? networks.small.evaluate(pos, accumulators, &caches.small)
: networks.big.evaluate(pos, accumulators, &caches.big);
Value nnue = (125 * psqt + 131 * positional) / 128;
// Re-evaluate the position when higher eval accuracy is worth the time spent
if (smallNet && (std::abs(nnue) < 236))
{
std::tie(psqt, positional) = networks.big.evaluate(pos, accumulators, &caches.big);
nnue = (125 * psqt + 131 * positional) / 128;
smallNet = false;
}
// Blend optimism and eval with nnue complexity
int nnueComplexity = std::abs(psqt - positional);
optimism += optimism * nnueComplexity / 468;
nnue -= nnue * nnueComplexity / 18000;
int material = 535 * pos.count<PAWN>() + pos.non_pawn_material();
int v = (nnue * (77777 + material) + optimism * (7777 + material)) / 77777;
// Damp down the evaluation linearly when shuffling
v -= v * pos.rule50_count() / 212;
// Guarantee evaluation does not hit the tablebase range
v = std::clamp(v, VALUE_TB_LOSS_IN_MAX_PLY + 1, VALUE_TB_WIN_IN_MAX_PLY - 1);
return v;
}
// Like evaluate(), but instead of returning a value, it returns
// a string (suitable for outputting to stdout) that contains the detailed
// descriptions and values of each evaluation term. Useful for debugging.
// Trace scores are from white's point of view
std::string Eval::trace(Position& pos, const Eval::NNUE::Networks& networks) {
if (pos.checkers())
return "Final evaluation: none (in check)";
auto accumulators = std::make_unique<Eval::NNUE::AccumulatorStack>();
auto caches = std::make_unique<Eval::NNUE::AccumulatorCaches>(networks);
std::stringstream ss;
ss << std::showpoint << std::noshowpos << std::fixed << std::setprecision(2);
ss << '\n' << NNUE::trace(pos, networks, *caches) << '\n';
ss << std::showpoint << std::showpos << std::fixed << std::setprecision(2) << std::setw(15);
auto [psqt, positional] = networks.big.evaluate(pos, *accumulators, &caches->big);
Value v = psqt + positional;
v = pos.side_to_move() == WHITE ? v : -v;
ss << "NNUE evaluation " << 0.01 * UCIEngine::to_cp(v, pos) << " (white side)\n";
v = evaluate(networks, pos, *accumulators, *caches, VALUE_ZERO);
v = pos.side_to_move() == WHITE ? v : -v;
ss << "Final evaluation " << 0.01 * UCIEngine::to_cp(v, pos) << " (white side)";
ss << " [with scaled NNUE, ...]";
ss << "\n";
return ss.str();
}
} // namespace Stockfish