opelly/Stockfish
1
0
Fork 0
mirror of https://github.com/opelly27/Stockfish.git synced 2026-05-20 08:37:44 +00:00
Code Issues Packages Projects Releases Wiki Activity

Refactor Network Usage

Browse Source 
Continuing from PR #4968, this update improves how Stockfish handles network
usage, making it easier to manage and modify networks in the future.

With the introduction of a dedicated Network class, creating networks has become
straightforward. See uci.cpp:
```cpp
NN::NetworkBig({EvalFileDefaultNameBig, "None", ""}, NN::embeddedNNUEBig)
```

The new `Network` encapsulates all network-related logic, significantly reducing
the complexity previously required to support multiple network types, such as
the distinction between small and big networks #4915.

Non-Regression STC:
https://tests.stockfishchess.org/tests/view/65edd26c0ec64f0526c43584
LLR: 2.94 (-2.94,2.94) <-1.75,0.25>
Total: 33760 W: 8887 L: 8661 D: 16212
Ptnml(0-2): 143, 3795, 8808, 3961, 173

Non-Regression SMP STC:
https://tests.stockfishchess.org/tests/view/65ed71970ec64f0526c42fdd
LLR: 2.96 (-2.94,2.94) <-1.75,0.25>
Total: 59088 W: 15121 L: 14931 D: 29036
Ptnml(0-2): 110, 6640, 15829, 6880, 85

Compiled with `make -j profile-build`
```
bash ./bench_parallel.sh ./stockfish ./stockfish-nnue 13 50

sf_base =  1568540 +/-   7637 (95%)
sf_test =  1573129 +/-   7301 (95%)
diff    =     4589 +/-   8720 (95%)
speedup = 0.29260% +/- 0.556% (95%)
```

Compiled with `make -j build`
```
bash ./bench_parallel.sh ./stockfish ./stockfish-nnue 13 50

sf_base =  1472653 +/-   7293 (95%)
sf_test =  1491928 +/-   7661 (95%)
diff    =    19275 +/-   7154 (95%)
speedup = 1.30886% +/- 0.486% (95%)
```

closes https://github.com/official-stockfish/Stockfish/pull/5100

No functional change
This commit is contained in:
Disservin
2024-03-09 14:42:37 +01:00
parent f072634e24
commit 1a26d698de
18 changed files with 948 additions and 826 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/nnue/evaluate_nnue.cpp
-488
View File
@@ -1,488 +0,0 @@
/*
Stockfish, a UCI chess playing engine derived from Glaurung 2.1
Copyright (C) 2004-2024 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/>.
*/
// Code for calculating NNUE evaluation function
#include "evaluate_nnue.h"
#include <cmath>
#include <cstdlib>
#include <cstring>
#include <fstream>
#include <iomanip>
#include <iostream>
#include <optional>
#include <sstream>
#include <string_view>
#include <type_traits>
#include "../evaluate.h"
#include "../misc.h"
#include "../position.h"
#include "../types.h"
#include "../uci.h"
#include "nnue_accumulator.h"
#include "nnue_common.h"
namespace Stockfish::Eval::NNUE {
// Input feature converter
LargePagePtr<FeatureTransformer<TransformedFeatureDimensionsBig, &StateInfo::accumulatorBig>>
featureTransformerBig;
LargePagePtr<FeatureTransformer<TransformedFeatureDimensionsSmall, &StateInfo::accumulatorSmall>>
featureTransformerSmall;
// Evaluation function
AlignedPtr<Network<TransformedFeatureDimensionsBig, L2Big, L3Big>> networkBig[LayerStacks];
AlignedPtr<Network<TransformedFeatureDimensionsSmall, L2Small, L3Small>> networkSmall[LayerStacks];
// Evaluation function file names
namespace Detail {
// Initialize the evaluation function parameters
template<typename T>
void initialize(AlignedPtr<T>& pointer) {
pointer.reset(reinterpret_cast<T*>(std_aligned_alloc(alignof(T), sizeof(T))));
std::memset(pointer.get(), 0, sizeof(T));
}
template<typename T>
void initialize(LargePagePtr<T>& pointer) {
static_assert(alignof(T) <= 4096,
"aligned_large_pages_alloc() may fail for such a big alignment requirement of T");
pointer.reset(reinterpret_cast<T*>(aligned_large_pages_alloc(sizeof(T))));
std::memset(pointer.get(), 0, sizeof(T));
}
// Read evaluation function parameters
template<typename T>
bool read_parameters(std::istream& stream, T& reference) {
std::uint32_t header;
header = read_little_endian<std::uint32_t>(stream);
if (!stream || header != T::get_hash_value())
return false;
return reference.read_parameters(stream);
}
// Write evaluation function parameters
template<typename T>
bool write_parameters(std::ostream& stream, const T& reference) {
write_little_endian<std::uint32_t>(stream, T::get_hash_value());
return reference.write_parameters(stream);
}
} // namespace Detail
// Initialize the evaluation function parameters
static void initialize(NetSize netSize) {
if (netSize == Small)
{
Detail::initialize(featureTransformerSmall);
for (std::size_t i = 0; i < LayerStacks; ++i)
Detail::initialize(networkSmall[i]);
}
else
{
Detail::initialize(featureTransformerBig);
for (std::size_t i = 0; i < LayerStacks; ++i)
Detail::initialize(networkBig[i]);
}
}
// Read network header
static bool read_header(std::istream& stream, std::uint32_t* hashValue, std::string* desc) {
std::uint32_t version, size;
version = read_little_endian<std::uint32_t>(stream);
*hashValue = read_little_endian<std::uint32_t>(stream);
size = read_little_endian<std::uint32_t>(stream);
if (!stream || version != Version)
return false;
desc->resize(size);
stream.read(&(*desc)[0], size);
return !stream.fail();
}
// Write network header
static bool write_header(std::ostream& stream, std::uint32_t hashValue, const std::string& desc) {
write_little_endian<std::uint32_t>(stream, Version);
write_little_endian<std::uint32_t>(stream, hashValue);
write_little_endian<std::uint32_t>(stream, std::uint32_t(desc.size()));
stream.write(&desc[0], desc.size());
return !stream.fail();
}
// Read network parameters
static bool read_parameters(std::istream& stream, NetSize netSize, std::string& netDescription) {
std::uint32_t hashValue;
if (!read_header(stream, &hashValue, &netDescription))
return false;
if (hashValue != HashValue[netSize])
return false;
if (netSize == Big && !Detail::read_parameters(stream, *featureTransformerBig))
return false;
if (netSize == Small && !Detail::read_parameters(stream, *featureTransformerSmall))
return false;
for (std::size_t i = 0; i < LayerStacks; ++i)
{
if (netSize == Big && !Detail::read_parameters(stream, *(networkBig[i])))
return false;
if (netSize == Small && !Detail::read_parameters(stream, *(networkSmall[i])))
return false;
}
return stream && stream.peek() == std::ios::traits_type::eof();
}
// Write network parameters
static bool
write_parameters(std::ostream& stream, NetSize netSize, const std::string& netDescription) {
if (!write_header(stream, HashValue[netSize], netDescription))
return false;
if (netSize == Big && !Detail::write_parameters(stream, *featureTransformerBig))
return false;
if (netSize == Small && !Detail::write_parameters(stream, *featureTransformerSmall))
return false;
for (std::size_t i = 0; i < LayerStacks; ++i)
{
if (netSize == Big && !Detail::write_parameters(stream, *(networkBig[i])))
return false;
if (netSize == Small && !Detail::write_parameters(stream, *(networkSmall[i])))
return false;
}
return bool(stream);
}
void hint_common_parent_position(const Position& pos) {
int simpleEvalAbs = std::abs(simple_eval(pos, pos.side_to_move()));
if (simpleEvalAbs > Eval::SmallNetThreshold)
featureTransformerSmall->hint_common_access(pos, simpleEvalAbs > Eval::PsqtOnlyThreshold);
else
featureTransformerBig->hint_common_access(pos, false);
}
// Evaluation function. Perform differential calculation.
template<NetSize Net_Size>
Value evaluate(const Position& pos, bool adjusted, int* complexity, bool psqtOnly) {
// We manually align the arrays on the stack because with gcc < 9.3
// overaligning stack variables with alignas() doesn't work correctly.
constexpr uint64_t alignment = CacheLineSize;
constexpr int delta = 24;
#if defined(ALIGNAS_ON_STACK_VARIABLES_BROKEN)
TransformedFeatureType transformedFeaturesUnaligned
[FeatureTransformer < Net_Size == Small ? TransformedFeatureDimensionsSmall
: TransformedFeatureDimensionsBig,
nullptr > ::BufferSize + alignment / sizeof(TransformedFeatureType)];
auto* transformedFeatures = align_ptr_up<alignment>(&transformedFeaturesUnaligned[0]);
#else
alignas(alignment) TransformedFeatureType
transformedFeatures[FeatureTransformer < Net_Size == Small ? TransformedFeatureDimensionsSmall
: TransformedFeatureDimensionsBig,
nullptr > ::BufferSize];
#endif
ASSERT_ALIGNED(transformedFeatures, alignment);
const int bucket = (pos.count<ALL_PIECES>() - 1) / 4;
const auto psqt =
Net_Size == Small
? featureTransformerSmall->transform(pos, transformedFeatures, bucket, psqtOnly)
: featureTransformerBig->transform(pos, transformedFeatures, bucket, psqtOnly);
const auto positional =
!psqtOnly ? (Net_Size == Small ? networkSmall[bucket]->propagate(transformedFeatures)
: networkBig[bucket]->propagate(transformedFeatures))
: 0;
if (complexity)
*complexity = !psqtOnly ? std::abs(psqt - positional) / OutputScale : 0;
// Give more value to positional evaluation when adjusted flag is set
if (adjusted)
return static_cast<Value>(((1024 - delta) * psqt + (1024 + delta) * positional)
/ (1024 * OutputScale));
else
return static_cast<Value>((psqt + positional) / OutputScale);
}
template Value evaluate<Big>(const Position& pos, bool adjusted, int* complexity, bool psqtOnly);
template Value evaluate<Small>(const Position& pos, bool adjusted, int* complexity, bool psqtOnly);
struct NnueEvalTrace {
static_assert(LayerStacks == PSQTBuckets);
Value psqt[LayerStacks];
Value positional[LayerStacks];
std::size_t correctBucket;
};
static NnueEvalTrace trace_evaluate(const Position& pos) {
// We manually align the arrays on the stack because with gcc < 9.3
// overaligning stack variables with alignas() doesn't work correctly.
constexpr uint64_t alignment = CacheLineSize;
#if defined(ALIGNAS_ON_STACK_VARIABLES_BROKEN)
TransformedFeatureType transformedFeaturesUnaligned
[FeatureTransformer<TransformedFeatureDimensionsBig, nullptr>::BufferSize
+ alignment / sizeof(TransformedFeatureType)];
auto* transformedFeatures = align_ptr_up<alignment>(&transformedFeaturesUnaligned[0]);
#else
alignas(alignment) TransformedFeatureType
transformedFeatures[FeatureTransformer<TransformedFeatureDimensionsBig, nullptr>::BufferSize];
#endif
ASSERT_ALIGNED(transformedFeatures, alignment);
NnueEvalTrace t{};
t.correctBucket = (pos.count<ALL_PIECES>() - 1) / 4;
for (IndexType bucket = 0; bucket < LayerStacks; ++bucket)
{
const auto materialist =
featureTransformerBig->transform(pos, transformedFeatures, bucket, false);
const auto positional = networkBig[bucket]->propagate(transformedFeatures);
t.psqt[bucket] = static_cast<Value>(materialist / OutputScale);
t.positional[bucket] = static_cast<Value>(positional / OutputScale);
}
return t;
}
constexpr std::string_view PieceToChar(" PNBRQK pnbrqk");
// Converts a Value into (centi)pawns and writes it in a buffer.
// The buffer must have capacity for at least 5 chars.
static void format_cp_compact(Value v, char* buffer) {
buffer[0] = (v < 0 ? '-' : v > 0 ? '+' : ' ');
int cp = std::abs(UCI::to_cp(v));
if (cp >= 10000)
{
buffer[1] = '0' + cp / 10000;
cp %= 10000;
buffer[2] = '0' + cp / 1000;
cp %= 1000;
buffer[3] = '0' + cp / 100;
buffer[4] = ' ';
}
else if (cp >= 1000)
{
buffer[1] = '0' + cp / 1000;
cp %= 1000;
buffer[2] = '0' + cp / 100;
cp %= 100;
buffer[3] = '.';
buffer[4] = '0' + cp / 10;
}
else
{
buffer[1] = '0' + cp / 100;
cp %= 100;
buffer[2] = '.';
buffer[3] = '0' + cp / 10;
cp %= 10;
buffer[4] = '0' + cp / 1;
}
}
// Converts a Value into pawns, always keeping two decimals
static void format_cp_aligned_dot(Value v, std::stringstream& stream) {
const double pawns = std::abs(0.01 * UCI::to_cp(v));
stream << (v < 0 ? '-'
: v > 0 ? '+'
: ' ')
<< std::setiosflags(std::ios::fixed) << std::setw(6) << std::setprecision(2) << pawns;
}
// Returns a string with the value of each piece on a board,
// and a table for (PSQT, Layers) values bucket by bucket.
std::string trace(Position& pos) {
std::stringstream ss;
char board[3 * 8 + 1][8 * 8 + 2];
std::memset(board, ' ', sizeof(board));
for (int row = 0; row < 3 * 8 + 1; ++row)
board[row][8 * 8 + 1] = '\0';
// A lambda to output one box of the board
auto writeSquare = [&board](File file, Rank rank, Piece pc, Value value) {
const int x = int(file) * 8;
const int y = (7 - int(rank)) * 3;
for (int i = 1; i < 8; ++i)
board[y][x + i] = board[y + 3][x + i] = '-';
for (int i = 1; i < 3; ++i)
board[y + i][x] = board[y + i][x + 8] = '|';
board[y][x] = board[y][x + 8] = board[y + 3][x + 8] = board[y + 3][x] = '+';
if (pc != NO_PIECE)
board[y + 1][x + 4] = PieceToChar[pc];
if (value != VALUE_NONE)
format_cp_compact(value, &board[y + 2][x + 2]);
};
// We estimate the value of each piece by doing a differential evaluation from
// the current base eval, simulating the removal of the piece from its square.
Value base = evaluate<NNUE::Big>(pos);
base = pos.side_to_move() == WHITE ? base : -base;
for (File f = FILE_A; f <= FILE_H; ++f)
for (Rank r = RANK_1; r <= RANK_8; ++r)
{
Square sq = make_square(f, r);
Piece pc = pos.piece_on(sq);
Value v = VALUE_NONE;
if (pc != NO_PIECE && type_of(pc) != KING)
{
auto st = pos.state();
pos.remove_piece(sq);
st->accumulatorBig.computed[WHITE] = st->accumulatorBig.computed[BLACK] =
st->accumulatorBig.computedPSQT[WHITE] = st->accumulatorBig.computedPSQT[BLACK] =
false;
Value eval = evaluate<NNUE::Big>(pos);
eval = pos.side_to_move() == WHITE ? eval : -eval;
v = base - eval;
pos.put_piece(pc, sq);
st->accumulatorBig.computed[WHITE] = st->accumulatorBig.computed[BLACK] =
st->accumulatorBig.computedPSQT[WHITE] = st->accumulatorBig.computedPSQT[BLACK] =
false;
}
writeSquare(f, r, pc, v);
}
ss << " NNUE derived piece values:\n";
for (int row = 0; row < 3 * 8 + 1; ++row)
ss << board[row] << '\n';
ss << '\n';
auto t = trace_evaluate(pos);
ss << " NNUE network contributions "
<< (pos.side_to_move() == WHITE ? "(White to move)" : "(Black to move)") << std::endl
<< "+------------+------------+------------+------------+\n"
<< "| Bucket | Material | Positional | Total |\n"
<< "| | (PSQT) | (Layers) | |\n"
<< "+------------+------------+------------+------------+\n";
for (std::size_t bucket = 0; bucket < LayerStacks; ++bucket)
{
ss << "| " << bucket << " ";
ss << " | ";
format_cp_aligned_dot(t.psqt[bucket], ss);
ss << " "
<< " | ";
format_cp_aligned_dot(t.positional[bucket], ss);
ss << " "
<< " | ";
format_cp_aligned_dot(t.psqt[bucket] + t.positional[bucket], ss);
ss << " "
<< " |";
if (bucket == t.correctBucket)
ss << " <-- this bucket is used";
ss << '\n';
}
ss << "+------------+------------+------------+------------+\n";
return ss.str();
}
// Load eval, from a file stream or a memory stream
std::optional<std::string> load_eval(std::istream& stream, NetSize netSize) {
initialize(netSize);
std::string netDescription;
return read_parameters(stream, netSize, netDescription) ? std::make_optional(netDescription)
: std::nullopt;
}
// Save eval, to a file stream or a memory stream
bool save_eval(std::ostream& stream,
NetSize netSize,
const std::string& name,
const std::string& netDescription) {
if (name.empty() || name == "None")
return false;
return write_parameters(stream, netSize, netDescription);
}
// Save eval, to a file given by its name
bool save_eval(const std::optional<std::string>& filename,
NetSize netSize,
const EvalFiles& evalFiles) {
std::string actualFilename;
std::string msg;
if (filename.has_value())
actualFilename = filename.value();
else
{
if (evalFiles.at(netSize).current
!= (netSize == Small ? EvalFileDefaultNameSmall : EvalFileDefaultNameBig))
{
msg = "Failed to export a net. "
"A non-embedded net can only be saved if the filename is specified";
sync_cout << msg << sync_endl;
return false;
}
actualFilename = (netSize == Small ? EvalFileDefaultNameSmall : EvalFileDefaultNameBig);
}
std::ofstream stream(actualFilename, std::ios_base::binary);
bool saved = save_eval(stream, netSize, evalFiles.at(netSize).current,
evalFiles.at(netSize).netDescription);
msg = saved ? "Network saved successfully to " + actualFilename : "Failed to export a net";
sync_cout << msg << sync_endl;
return saved;
}
} // namespace Stockfish::Eval::NNUE
src/nnue/evaluate_nnue.h
-89
View File
@@ -1,89 +0,0 @@
/*
Stockfish, a UCI chess playing engine derived from Glaurung 2.1
Copyright (C) 2004-2024 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/>.
*/
// header used in NNUE evaluation function
#ifndef NNUE_EVALUATE_NNUE_H_INCLUDED
#define NNUE_EVALUATE_NNUE_H_INCLUDED
#include <cstdint>
#include <iosfwd>
#include <memory>
#include <optional>
#include <string>
#include "../evaluate.h"
#include "../misc.h"
#include "../types.h"
#include "nnue_architecture.h"
#include "nnue_feature_transformer.h"
namespace Stockfish {
class Position;
}
namespace Stockfish::Eval::NNUE {
// Hash value of evaluation function structure
constexpr std::uint32_t HashValue[2] = {
FeatureTransformer<TransformedFeatureDimensionsBig, nullptr>::get_hash_value()
^ Network<TransformedFeatureDimensionsBig, L2Big, L3Big>::get_hash_value(),
FeatureTransformer<TransformedFeatureDimensionsSmall, nullptr>::get_hash_value()
^ Network<TransformedFeatureDimensionsSmall, L2Small, L3Small>::get_hash_value()};
// Deleter for automating release of memory area
template<typename T>
struct AlignedDeleter {
void operator()(T* ptr) const {
ptr->~T();
std_aligned_free(ptr);
}
};
template<typename T>
struct LargePageDeleter {
void operator()(T* ptr) const {
ptr->~T();
aligned_large_pages_free(ptr);
}
};
template<typename T>
using AlignedPtr = std::unique_ptr<T, AlignedDeleter<T>>;
template<typename T>
using LargePagePtr = std::unique_ptr<T, LargePageDeleter<T>>;
std::string trace(Position& pos);
template<NetSize Net_Size>
Value evaluate(const Position& pos,
bool adjusted = false,
int* complexity = nullptr,
bool psqtOnly = false);
void hint_common_parent_position(const Position& pos);
std::optional<std::string> load_eval(std::istream& stream, NetSize netSize);
bool save_eval(std::ostream& stream,
NetSize netSize,
const std::string& name,
const std::string& netDescription);
bool save_eval(const std::optional<std::string>& filename, NetSize netSize, const EvalFiles&);
} // namespace Stockfish::Eval::NNUE
#endif // #ifndef NNUE_EVALUATE_NNUE_H_INCLUDED
src/nnue/network.cpp
+422
View File
@@ -0,0 +1,422 @@
/*
Stockfish, a UCI chess playing engine derived from Glaurung 2.1
Copyright (C) 2004-2024 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 "network.h"
#include <cmath>
#include <cstdlib>
#include <cstring>
#include <fstream>
#include <iostream>
#include <optional>
#include <type_traits>
#include <vector>
#include "../evaluate.h"
#include "../incbin/incbin.h"
#include "../misc.h"
#include "../position.h"
#include "../types.h"
#include "nnue_architecture.h"
#include "nnue_common.h"
#include "nnue_misc.h"
namespace {
// Macro to embed the default efficiently updatable neural network (NNUE) file
// data in the engine binary (using incbin.h, by Dale Weiler).
// This macro invocation will declare the following three variables
// const unsigned char gEmbeddedNNUEData[]; // a pointer to the embedded data
// const unsigned char *const gEmbeddedNNUEEnd; // a marker to the end
// const unsigned int gEmbeddedNNUESize; // the size of the embedded file
// Note that this does not work in Microsoft Visual Studio.
#if !defined(_MSC_VER) && !defined(NNUE_EMBEDDING_OFF)
INCBIN(EmbeddedNNUEBig, EvalFileDefaultNameBig);
INCBIN(EmbeddedNNUESmall, EvalFileDefaultNameSmall);
#else
const unsigned char gEmbeddedNNUEBigData[1] = {0x0};
const unsigned char* const gEmbeddedNNUEBigEnd = &gEmbeddedNNUEBigData[1];
const unsigned int gEmbeddedNNUEBigSize = 1;
const unsigned char gEmbeddedNNUESmallData[1] = {0x0};
const unsigned char* const gEmbeddedNNUESmallEnd = &gEmbeddedNNUESmallData[1];
const unsigned int gEmbeddedNNUESmallSize = 1;
#endif
}
namespace Stockfish::Eval::NNUE {
const EmbeddedNNUE embeddedNNUEBig(gEmbeddedNNUEBigData, gEmbeddedNNUEBigEnd, gEmbeddedNNUEBigSize);
const EmbeddedNNUE
embeddedNNUESmall(gEmbeddedNNUESmallData, gEmbeddedNNUESmallEnd, gEmbeddedNNUESmallSize);
namespace Detail {
// Initialize the evaluation function parameters
template<typename T>
void initialize(AlignedPtr<T>& pointer) {
pointer.reset(reinterpret_cast<T*>(std_aligned_alloc(alignof(T), sizeof(T))));
std::memset(pointer.get(), 0, sizeof(T));
}
template<typename T>
void initialize(LargePagePtr<T>& pointer) {
static_assert(alignof(T) <= 4096,
"aligned_large_pages_alloc() may fail for such a big alignment requirement of T");
pointer.reset(reinterpret_cast<T*>(aligned_large_pages_alloc(sizeof(T))));
std::memset(pointer.get(), 0, sizeof(T));
}
// Read evaluation function parameters
template<typename T>
bool read_parameters(std::istream& stream, T& reference) {
std::uint32_t header;
header = read_little_endian<std::uint32_t>(stream);
if (!stream || header != T::get_hash_value())
return false;
return reference.read_parameters(stream);
}
// Write evaluation function parameters
template<typename T>
bool write_parameters(std::ostream& stream, const T& reference) {
write_little_endian<std::uint32_t>(stream, T::get_hash_value());
return reference.write_parameters(stream);
}
} // namespace Detail
template<typename Arch, typename Transformer>
void Network<Arch, Transformer>::load(const std::string& rootDirectory, std::string evalfilePath) {
#if defined(DEFAULT_NNUE_DIRECTORY)
std::vector<std::string> dirs = {"<internal>", "", rootDirectory,
stringify(DEFAULT_NNUE_DIRECTORY)};
#else
std::vector<std::string> dirs = {"<internal>", "", rootDirectory};
#endif
if (evalfilePath.empty())
evalfilePath = evalFile.defaultName;
for (const auto& directory : dirs)
{
if (evalFile.current != evalfilePath)
{
if (directory != "<internal>")
{
load_user_net(directory, evalfilePath);
}
if (directory == "<internal>" && evalfilePath == evalFile.defaultName)
{
load_internal();
}
}
}
}
template<typename Arch, typename Transformer>
bool Network<Arch, Transformer>::save(const std::optional<std::string>& filename) const {
std::string actualFilename;
std::string msg;
if (filename.has_value())
actualFilename = filename.value();
else
{
if (evalFile.current != evalFile.defaultName)
{
msg = "Failed to export a net. "
"A non-embedded net can only be saved if the filename is specified";
sync_cout << msg << sync_endl;
return false;
}
actualFilename = evalFile.defaultName;
}
std::ofstream stream(actualFilename, std::ios_base::binary);
bool saved = save(stream, evalFile.current, evalFile.netDescription);
msg = saved ? "Network saved successfully to " + actualFilename : "Failed to export a net";
sync_cout << msg << sync_endl;
return saved;
}
template<typename Arch, typename Transformer>
Value Network<Arch, Transformer>::evaluate(const Position& pos,
bool adjusted,
int* complexity,
bool psqtOnly) const {
// We manually align the arrays on the stack because with gcc < 9.3
// overaligning stack variables with alignas() doesn't work correctly.
constexpr uint64_t alignment = CacheLineSize;
constexpr int delta = 24;
#if defined(ALIGNAS_ON_STACK_VARIABLES_BROKEN)
TransformedFeatureType transformedFeaturesUnaligned
[FeatureTransformer<Arch::TransformedFeatureDimensions, nullptr>::BufferSize
+ alignment / sizeof(TransformedFeatureType)];
auto* transformedFeatures = align_ptr_up<alignment>(&transformedFeaturesUnaligned[0]);
#else
alignas(alignment) TransformedFeatureType transformedFeatures
[FeatureTransformer<Arch::TransformedFeatureDimensions, nullptr>::BufferSize];
#endif
ASSERT_ALIGNED(transformedFeatures, alignment);
const int bucket = (pos.count<ALL_PIECES>() - 1) / 4;
const auto psqt = featureTransformer->transform(pos, transformedFeatures, bucket, psqtOnly);
const auto positional = !psqtOnly ? (network[bucket]->propagate(transformedFeatures)) : 0;
if (complexity)
*complexity = !psqtOnly ? std::abs(psqt - positional) / OutputScale : 0;
// Give more value to positional evaluation when adjusted flag is set
if (adjusted)
return static_cast<Value>(((1024 - delta) * psqt + (1024 + delta) * positional)
/ (1024 * OutputScale));
else
return static_cast<Value>((psqt + positional) / OutputScale);
}
template<typename Arch, typename Transformer>
void Network<Arch, Transformer>::verify(std::string evalfilePath) const {
if (evalfilePath.empty())
evalfilePath = evalFile.defaultName;
if (evalFile.current != evalfilePath)
{
std::string msg1 =
"Network evaluation parameters compatible with the engine must be available.";
std::string msg2 = "The network file " + evalfilePath + " was not loaded successfully.";
std::string msg3 = "The UCI option EvalFile might need to specify the full path, "
"including the directory name, to the network file.";
std::string msg4 = "The default net can be downloaded from: "
"https://tests.stockfishchess.org/api/nn/"
+ evalFile.defaultName;
std::string msg5 = "The engine will be terminated now.";
sync_cout << "info string ERROR: " << msg1 << sync_endl;
sync_cout << "info string ERROR: " << msg2 << sync_endl;
sync_cout << "info string ERROR: " << msg3 << sync_endl;
sync_cout << "info string ERROR: " << msg4 << sync_endl;
sync_cout << "info string ERROR: " << msg5 << sync_endl;
exit(EXIT_FAILURE);
}
sync_cout << "info string NNUE evaluation using " << evalfilePath << sync_endl;
}
template<typename Arch, typename Transformer>
void Network<Arch, Transformer>::hint_common_access(const Position& pos, bool psqtOnl) const {
featureTransformer->hint_common_access(pos, psqtOnl);
}
template<typename Arch, typename Transformer>
NnueEvalTrace Network<Arch, Transformer>::trace_evaluate(const Position& pos) const {
// We manually align the arrays on the stack because with gcc < 9.3
// overaligning stack variables with alignas() doesn't work correctly.
constexpr uint64_t alignment = CacheLineSize;
#if defined(ALIGNAS_ON_STACK_VARIABLES_BROKEN)
TransformedFeatureType transformedFeaturesUnaligned
[FeatureTransformer<Arch::TransformedFeatureDimensions, nullptr>::BufferSize
+ alignment / sizeof(TransformedFeatureType)];
auto* transformedFeatures = align_ptr_up<alignment>(&transformedFeaturesUnaligned[0]);
#else
alignas(alignment) TransformedFeatureType transformedFeatures
[FeatureTransformer<Arch::TransformedFeatureDimensions, nullptr>::BufferSize];
#endif
ASSERT_ALIGNED(transformedFeatures, alignment);
NnueEvalTrace t{};
t.correctBucket = (pos.count<ALL_PIECES>() - 1) / 4;
for (IndexType bucket = 0; bucket < LayerStacks; ++bucket)
{
const auto materialist =
featureTransformer->transform(pos, transformedFeatures, bucket, false);
const auto positional = network[bucket]->propagate(transformedFeatures);
t.psqt[bucket] = static_cast<Value>(materialist / OutputScale);
t.positional[bucket] = static_cast<Value>(positional / OutputScale);
}
return t;
}
template<typename Arch, typename Transformer>
void Network<Arch, Transformer>::load_user_net(const std::string& dir,
const std::string& evalfilePath) {
std::ifstream stream(dir + evalfilePath, std::ios::binary);
auto description = load(stream);
if (description.has_value())
{
evalFile.current = evalfilePath;
evalFile.netDescription = description.value();
}
}
template<typename Arch, typename Transformer>
void Network<Arch, Transformer>::load_internal() {
// C++ way to prepare a buffer for a memory stream
class MemoryBuffer: public std::basic_streambuf<char> {
public:
MemoryBuffer(char* p, size_t n) {
setg(p, p, p + n);
setp(p, p + n);
}
};
MemoryBuffer buffer(const_cast<char*>(reinterpret_cast<const char*>(embedded.data)),
size_t(embedded.size));
std::istream stream(&buffer);
auto description = load(stream);
if (description.has_value())
{
evalFile.current = evalFile.defaultName;
evalFile.netDescription = description.value();
}
}
template<typename Arch, typename Transformer>
void Network<Arch, Transformer>::initialize() {
Detail::initialize(featureTransformer);
for (std::size_t i = 0; i < LayerStacks; ++i)
Detail::initialize(network[i]);
}
template<typename Arch, typename Transformer>
bool Network<Arch, Transformer>::save(std::ostream& stream,
const std::string& name,
const std::string& netDescription) const {
if (name.empty() || name == "None")
return false;
return write_parameters(stream, netDescription);
}
template<typename Arch, typename Transformer>
std::optional<std::string> Network<Arch, Transformer>::load(std::istream& stream) {
initialize();
std::string description;
return read_parameters(stream, description) ? std::make_optional(description) : std::nullopt;
}
// Read network header
template<typename Arch, typename Transformer>
bool Network<Arch, Transformer>::read_header(std::istream& stream,
std::uint32_t* hashValue,
std::string* desc) const {
std::uint32_t version, size;
version = read_little_endian<std::uint32_t>(stream);
*hashValue = read_little_endian<std::uint32_t>(stream);
size = read_little_endian<std::uint32_t>(stream);
if (!stream || version != Version)
return false;
desc->resize(size);
stream.read(&(*desc)[0], size);
return !stream.fail();
}
// Write network header
template<typename Arch, typename Transformer>
bool Network<Arch, Transformer>::write_header(std::ostream& stream,
std::uint32_t hashValue,
const std::string& desc) const {
write_little_endian<std::uint32_t>(stream, Version);
write_little_endian<std::uint32_t>(stream, hashValue);
write_little_endian<std::uint32_t>(stream, std::uint32_t(desc.size()));
stream.write(&desc[0], desc.size());
return !stream.fail();
}
template<typename Arch, typename Transformer>
bool Network<Arch, Transformer>::read_parameters(std::istream& stream,
std::string& netDescription) const {
std::uint32_t hashValue;
if (!read_header(stream, &hashValue, &netDescription))
return false;
if (hashValue != Network::hash)
return false;
if (!Detail::read_parameters(stream, *featureTransformer))
return false;
for (std::size_t i = 0; i < LayerStacks; ++i)
{
if (!Detail::read_parameters(stream, *(network[i])))
return false;
}
return stream && stream.peek() == std::ios::traits_type::eof();
}
template<typename Arch, typename Transformer>
bool Network<Arch, Transformer>::write_parameters(std::ostream& stream,
const std::string& netDescription) const {
if (!write_header(stream, Network::hash, netDescription))
return false;
if (!Detail::write_parameters(stream, *featureTransformer))
return false;
for (std::size_t i = 0; i < LayerStacks; ++i)
{
if (!Detail::write_parameters(stream, *(network[i])))
return false;
}
return bool(stream);
}
// Explicit template instantiation
template class Network<
NetworkArchitecture<TransformedFeatureDimensionsBig, L2Big, L3Big>,
FeatureTransformer<TransformedFeatureDimensionsBig, &StateInfo::accumulatorBig>>;
template class Network<
NetworkArchitecture<TransformedFeatureDimensionsSmall, L2Small, L3Small>,
FeatureTransformer<TransformedFeatureDimensionsSmall, &StateInfo::accumulatorSmall>>;
} // namespace Stockfish::Eval::NNUE
src/nnue/network.h
+128
View File
@@ -0,0 +1,128 @@
/*
Stockfish, a UCI chess playing engine derived from Glaurung 2.1
Copyright (C) 2004-2024 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/>.
*/
#ifndef NETWORK_H_INCLUDED
#define NETWORK_H_INCLUDED
#include <cstdint>
#include <iostream>
#include <optional>
#include <string>
#include <utility>
#include "../misc.h"
#include "../position.h"
#include "../types.h"
#include "nnue_architecture.h"
#include "nnue_feature_transformer.h"
#include "nnue_misc.h"
namespace Stockfish::Eval::NNUE {
struct EmbeddedNNUE {
EmbeddedNNUE(const unsigned char* embeddedData,
const unsigned char* embeddedEnd,
const unsigned int embeddedSize) :
data(embeddedData),
end(embeddedEnd),
size(embeddedSize) {}
const unsigned char* data;
const unsigned char* end;
const unsigned int size;
};
extern const EmbeddedNNUE embeddedNNUEBig;
extern const EmbeddedNNUE embeddedNNUESmall;
template<typename Arch, typename Transformer>
class Network {
public:
Network(EvalFile file, EmbeddedNNUE embeddedEval) :
evalFile(file),
embedded(embeddedEval) {}
void load(const std::string& rootDirectory, std::string evalfilePath);
bool save(const std::optional<std::string>& filename) const;
Value evaluate(const Position& pos,
bool adjusted = false,
int* complexity = nullptr,
bool psqtOnly = false) const;
void hint_common_access(const Position& pos, bool psqtOnl) const;
void verify(std::string evalfilePath) const;
NnueEvalTrace trace_evaluate(const Position& pos) const;
private:
void load_user_net(const std::string&, const std::string&);
void load_internal();
void initialize();
bool save(std::ostream&, const std::string&, const std::string&) const;
std::optional<std::string> load(std::istream&);
bool read_header(std::istream&, std::uint32_t*, std::string*) const;
bool write_header(std::ostream&, std::uint32_t, const std::string&) const;
bool read_parameters(std::istream&, std::string&) const;
bool write_parameters(std::ostream&, const std::string&) const;
// Input feature converter
LargePagePtr<Transformer> featureTransformer;
// Evaluation function
AlignedPtr<Arch> network[LayerStacks];
EvalFile evalFile;
EmbeddedNNUE embedded;
// Hash value of evaluation function structure
static constexpr std::uint32_t hash = Transformer::get_hash_value() ^ Arch::get_hash_value();
};
// Definitions of the network types
using SmallFeatureTransformer =
FeatureTransformer<TransformedFeatureDimensionsSmall, &StateInfo::accumulatorSmall>;
using SmallNetworkArchitecture =
NetworkArchitecture<TransformedFeatureDimensionsSmall, L2Small, L3Small>;
using BigFeatureTransformer =
FeatureTransformer<TransformedFeatureDimensionsBig, &StateInfo::accumulatorBig>;
using BigNetworkArchitecture = NetworkArchitecture<TransformedFeatureDimensionsBig, L2Big, L3Big>;
using NetworkBig = Network<BigNetworkArchitecture, BigFeatureTransformer>;
using NetworkSmall = Network<SmallNetworkArchitecture, SmallFeatureTransformer>;
struct Networks {
Networks(NetworkBig&& nB, NetworkSmall&& nS) :
big(std::move(nB)),
small(std::move(nS)) {}
NetworkBig big;
NetworkSmall small;
};
} // namespace Stockfish
#endif
src/nnue/nnue_architecture.h
+1 -6
View File
@@ -37,11 +37,6 @@ namespace Stockfish::Eval::NNUE {
// Input features used in evaluation function
using FeatureSet = Features::HalfKAv2_hm;
enum NetSize : int {
Big,
Small
};
// Number of input feature dimensions after conversion
constexpr IndexType TransformedFeatureDimensionsBig = 2560;
constexpr int L2Big = 15;
@@ -55,7 +50,7 @@ constexpr IndexType PSQTBuckets = 8;
constexpr IndexType LayerStacks = 8;
template<IndexType L1, int L2, int L3>
struct Network {
struct NetworkArchitecture {
static constexpr IndexType TransformedFeatureDimensions = L1;
static constexpr int FC_0_OUTPUTS = L2;
static constexpr int FC_1_OUTPUTS = L3;
src/nnue/nnue_misc.cpp
+202
View File
@@ -0,0 +1,202 @@
/*
Stockfish, a UCI chess playing engine derived from Glaurung 2.1
Copyright (C) 2004-2024 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/>.
*/
// Code for calculating NNUE evaluation function
#include "nnue_misc.h"
#include <cmath>
#include <cstdlib>
#include <cstring>
#include <iomanip>
#include <iosfwd>
#include <iostream>
#include <sstream>
#include <string_view>
#include "../evaluate.h"
#include "../position.h"
#include "../types.h"
#include "../uci.h"
#include "network.h"
#include "nnue_accumulator.h"
namespace Stockfish::Eval::NNUE {
constexpr std::string_view PieceToChar(" PNBRQK pnbrqk");
void hint_common_parent_position(const Position& pos, const Networks& networks) {
int simpleEvalAbs = std::abs(simple_eval(pos, pos.side_to_move()));
if (simpleEvalAbs > Eval::SmallNetThreshold)
networks.small.hint_common_access(pos, simpleEvalAbs > Eval::PsqtOnlyThreshold);
else
networks.big.hint_common_access(pos, false);
}
// Converts a Value into (centi)pawns and writes it in a buffer.
// The buffer must have capacity for at least 5 chars.
static void format_cp_compact(Value v, char* buffer) {
buffer[0] = (v < 0 ? '-' : v > 0 ? '+' : ' ');
int cp = std::abs(UCI::to_cp(v));
if (cp >= 10000)
{
buffer[1] = '0' + cp / 10000;
cp %= 10000;
buffer[2] = '0' + cp / 1000;
cp %= 1000;
buffer[3] = '0' + cp / 100;
buffer[4] = ' ';
}
else if (cp >= 1000)
{
buffer[1] = '0' + cp / 1000;
cp %= 1000;
buffer[2] = '0' + cp / 100;
cp %= 100;
buffer[3] = '.';
buffer[4] = '0' + cp / 10;
}
else
{
buffer[1] = '0' + cp / 100;
cp %= 100;
buffer[2] = '.';
buffer[3] = '0' + cp / 10;
cp %= 10;
buffer[4] = '0' + cp / 1;
}
}
// Converts a Value into pawns, always keeping two decimals
static void format_cp_aligned_dot(Value v, std::stringstream& stream) {
const double pawns = std::abs(0.01 * UCI::to_cp(v));
stream << (v < 0 ? '-'
: v > 0 ? '+'
: ' ')
<< std::setiosflags(std::ios::fixed) << std::setw(6) << std::setprecision(2) << pawns;
}
// Returns a string with the value of each piece on a board,
// and a table for (PSQT, Layers) values bucket by bucket.
std::string trace(Position& pos, const Eval::NNUE::Networks& networks) {
std::stringstream ss;
char board[3 * 8 + 1][8 * 8 + 2];
std::memset(board, ' ', sizeof(board));
for (int row = 0; row < 3 * 8 + 1; ++row)
board[row][8 * 8 + 1] = '\0';
// A lambda to output one box of the board
auto writeSquare = [&board](File file, Rank rank, Piece pc, Value value) {
const int x = int(file) * 8;
const int y = (7 - int(rank)) * 3;
for (int i = 1; i < 8; ++i)
board[y][x + i] = board[y + 3][x + i] = '-';
for (int i = 1; i < 3; ++i)
board[y + i][x] = board[y + i][x + 8] = '|';
board[y][x] = board[y][x + 8] = board[y + 3][x + 8] = board[y + 3][x] = '+';
if (pc != NO_PIECE)
board[y + 1][x + 4] = PieceToChar[pc];
if (value != VALUE_NONE)
format_cp_compact(value, &board[y + 2][x + 2]);
};
// We estimate the value of each piece by doing a differential evaluation from
// the current base eval, simulating the removal of the piece from its square.
Value base = networks.big.evaluate(pos);
base = pos.side_to_move() == WHITE ? base : -base;
for (File f = FILE_A; f <= FILE_H; ++f)
for (Rank r = RANK_1; r <= RANK_8; ++r)
{
Square sq = make_square(f, r);
Piece pc = pos.piece_on(sq);
Value v = VALUE_NONE;
if (pc != NO_PIECE && type_of(pc) != KING)
{
auto st = pos.state();
pos.remove_piece(sq);
st->accumulatorBig.computed[WHITE] = st->accumulatorBig.computed[BLACK] =
st->accumulatorBig.computedPSQT[WHITE] = st->accumulatorBig.computedPSQT[BLACK] =
false;
Value eval = networks.big.evaluate(pos);
eval = pos.side_to_move() == WHITE ? eval : -eval;
v = base - eval;
pos.put_piece(pc, sq);
st->accumulatorBig.computed[WHITE] = st->accumulatorBig.computed[BLACK] =
st->accumulatorBig.computedPSQT[WHITE] = st->accumulatorBig.computedPSQT[BLACK] =
false;
}
writeSquare(f, r, pc, v);
}
ss << " NNUE derived piece values:\n";
for (int row = 0; row < 3 * 8 + 1; ++row)
ss << board[row] << '\n';
ss << '\n';
auto t = networks.big.trace_evaluate(pos);
ss << " NNUE network contributions "
<< (pos.side_to_move() == WHITE ? "(White to move)" : "(Black to move)") << std::endl
<< "+------------+------------+------------+------------+\n"
<< "| Bucket | Material | Positional | Total |\n"
<< "| | (PSQT) | (Layers) | |\n"
<< "+------------+------------+------------+------------+\n";
for (std::size_t bucket = 0; bucket < LayerStacks; ++bucket)
{
ss << "| " << bucket << " ";
ss << " | ";
format_cp_aligned_dot(t.psqt[bucket], ss);
ss << " "
<< " | ";
format_cp_aligned_dot(t.positional[bucket], ss);
ss << " "
<< " | ";
format_cp_aligned_dot(t.psqt[bucket] + t.positional[bucket], ss);
ss << " "
<< " |";
if (bucket == t.correctBucket)
ss << " <-- this bucket is used";
ss << '\n';
}
ss << "+------------+------------+------------+------------+\n";
return ss.str();
}
} // namespace Stockfish::Eval::NNUE
src/nnue/nnue_misc.h
+63
View File
@@ -0,0 +1,63 @@
/*
Stockfish, a UCI chess playing engine derived from Glaurung 2.1
Copyright (C) 2004-2024 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/>.
*/
#ifndef NNUE_MISC_H_INCLUDED
#define NNUE_MISC_H_INCLUDED
#include <cstddef>
#include <string>
#include "../types.h"
#include "nnue_architecture.h"
namespace Stockfish {
class Position;
namespace Eval::NNUE {
struct EvalFile {
// Default net name, will use one of the EvalFileDefaultName* macros defined
// in evaluate.h
std::string defaultName;
// Selected net name, either via uci option or default
std::string current;
// Net description extracted from the net file
std::string netDescription;
};
struct NnueEvalTrace {
static_assert(LayerStacks == PSQTBuckets);
Value psqt[LayerStacks];
Value positional[LayerStacks];
std::size_t correctBucket;
};
struct Networks;
std::string trace(Position& pos, const Networks& networks);
void hint_common_parent_position(const Position& pos, const Networks& networks);
} // namespace Stockfish::Eval::NNUE
} // namespace Stockfish
#endif // #ifndef NNUE_MISC_H_INCLUDED